Face Recognition in Challenging Situations - Publication

Face Recognition in
Challenging Situations

Eilidh Clare Noyes

Doctor of Philosophy
University of York
Psychology

May 2016

Abstract

A great deal of previous research has demonstrated that face recognition is unreliable for
unfamiliar faces and reliable for familiar faces. However, such findings typically came
from tasks that used ‘cooperative’ images, where there was no deliberate attempt to
alter apparent identity. In applied settings, images are often far more challenging in
nature. For example multiple images of the same identity may appear to be different
identities, due to either incidental changes in appearance (such as age or style related
change, or differences in images capture) or deliberate changes (evading own identity
through disguise). At the same time, images of different identities may look like the same
person, due to either incidental changes (natural similarities in appearance), or deliberate
changes (attempts to impersonate someone else, such as in the case of identity fraud).
Thus, past studies may have underestimated the applied problem. In this thesis I examine
face recognition performance for these challenging image scenarios and test whether the
familiarity advantage extends to these situations. I found that face recognition was
indeed even poorer for challenging images than previously found using cooperative
images. Familiar viewers were still better than unfamiliar viewers, yet familiarity did not
bring performance to ceiling level for challenging images as it had done in the cooperative
tasks in the past. I investigated several ways of improving performance, including image
manipulations, exploiting perceptual constancy, crowd analysis of identity judgments, and
viewing by super-recognisers. This thesis provides interesting insights into theory
regarding what it is that familiar viewers are learning when they are becoming familiar
with a face. It also has important practical implications; both for improving performance
in challenging situations and for understanding deliberate disguise.

2

Table of Contents

Abstract ...................................................................... 2
Table of Contents ....................................................... 3
List of Tables ............................................................... 6
List of Figures .............................................................. 7
Acknowledgements .................................................. 15
Declaration ............................................................... 16
Chapter 1 – General Introduction ............................. 17
1.1 Introduction: Why Face Recognition is Important ................................................. 17
1.2 Familiarity & Face-Matching .................................................................................. 21
1.3 What Information is used to Recognise a Face? .................................................... 30
1.4 Learning Variability ................................................................................................ 36
1.5 Face Recognition in Challenging Situations ........................................................... 40
1.6 Overview of Current Work ..................................................................................... 44

Chapter 2 – Familiarity & Challenging Faces ............ 48
2.1 Chapter Summary .................................................................................................. 48
2.2 Introduction ........................................................................................................... 48
2.3 Celebrity Faces & Celebrity Lookalikes .................................................................. 54
2.4 Experiment 1: Lookalike Task ................................................................................. 55
2.5 Experiment 2: Mid Pixelation ................................................................................. 65

3

2.6 Experiment 3: Coarse Pixelation ............................................................................ 74
2.7 Between Experiments Analysis .............................................................................. 81
2.8 General Discussion ................................................................................................. 82

Chapter 3 – Improving Performance ........................ 86
3.1 Chapter Summary .................................................................................................. 86
3.2 Introduction ........................................................................................................... 86
3.3 Experiment 4: Blurring Pixelated Images ............................................................... 91
3.4 Experiment 5: Crowd Analysis ............................................................................... 97
3.5 Experiment 6: Observer Factors .......................................................................... 107
3.6 General Discussion ............................................................................................... 117

Chapter 4 – Changing Camera-to-Subject Distance 120
4.1 Chapter Summary ................................................................................................ 120
4.2 Introduction ......................................................................................................... 120
4.3 Experiment 7: Facial Configuration Measurements ............................................. 124
4.4 Experiment 8: Face-Matching & Camera-to-Subject Distance ............................. 127
4.5 Experiment 9 - Perceptual Constancy for Face Shape ......................................... 134
4.6 General Discussion ............................................................................................... 140

Chapter 5 – Matching Disguised Faces ................... 144
5.1 Chapter Summary ................................................................................................ 144
5.2 Introduction ......................................................................................................... 144
5.3 Existing Disguise Face Databases ......................................................................... 152
5.4 FAÇADE Database ................................................................................................ 156
5.5 Experiment 10: Unfamiliar Viewers ..................................................................... 166
5.6 Experiment 11: Unfamiliar (Informed) Viewers ................................................... 171

4

5.7 Experiment 12: Familiar Viewers ......................................................................... 176
5.8 General Discussion ............................................................................................... 184

Chapter 6 – Understanding Disguise ...................... 191
6.1 Chapter Summary ................................................................................................ 191
6.2 Introduction ......................................................................................................... 192
6.3 How do People Disguise Themselves? ................................................................. 199
6.4 Can Viewers Predict by Eye which Disguises will be Effective? ........................... 205
6.5 What do Viewers Believe Makes for an Effective Disguise? ................................ 208
6.6 Experiment 13 - Do Social Inferences Change for Disguise? ................................ 213
6.7 General Discussion ............................................................................................... 221

Chapter 7 – General Discussion .............................. 224
7.1 Overview of Findings ........................................................................................... 224
7.2 Relation to Previous Research ............................................................................. 228
7.3 Theoretical Implications ...................................................................................... 231
7.4 Practical Implications ........................................................................................... 235
7.5 Future Directions ................................................................................................. 237

References .............................................................. 241

5

List of Tables

Table 3.1 Crowd analysis results broken down by trial types (same identity, different
identity)……….……………………………………………………………………………………………………………….104
Table 3.2 Crowd analysis results broken down by trial types (same identity, different
identity)………………………………………………………………………………………………………………….….…106
Table 3.3 Performance accuracy broken down by viewer group and trial type……………..115
Table 3.4 Crowd analysis results broken down by trial types (same identity, different
identity) for the mid pixelation (control) Experiment 2…………………………………………………116
Table 3.5 Crowd analysis results broken down by trial types (same identity, different
identity) for the SRs.…………………………………………………………………………………………….……….116
Table 4.1 Table showing mean measurements for each photograph condition. EN stands
for ear to nose measurement, and NM represents nose to mouth. The letters following
denote the side of the image which the measurement was taken for, L = left, R = right & C
= centre. Average measurements are calculated for week 1 (Avg1) and week 2 (Avg2) at
both near (AvgN) and far (AvgF) distances………………………………….………………………………..127

Table 6.1 Social inference comparisons for impersonation similar images……………………219
Table 6.2. Impersonation random disguises, means and median results for each
analysis…………………………………………………………………………………………………………………………220

6

List of Figures

Figure 1.1 Image showing the face of the real suspect, Hussain Oman on the left and the
face of Jean Charles de Menezes on the right…………………………………………………………………18
Figure 1.2 Face-matching performance of passport officers in the study by White et al.
(2014). Some officers perform with very high performance regardless of their
employment duration..…………………………………………………………………………………………………..24
Figure 1.3 Example of line stimuli taken from Bruce et al. (1999) line up task. The correct
match of the target face is face number 3.……………………………………………………………………..25
Figure 1.4 Examples of face pairs taken from the GFMT (Burton et al., 2010). Top row
show different identity pairs, bottom row are same identity pairs…………………………………27
Figure 1.5 Examples of face pairs in the matching task involving poor quality CCTV footage
in the experiments conducted by Burton et al. (1999). Familiar viewers performed with
high accuracy when matching the face pairs...…………………………………………………………..……28
Figure 1.6 Examples of full-face image on the left was to be matched with either of the
internal face images on the right. Familiarity aided matching accuracy on this task……….29
Figure 1.7 Example of test stimuli used by Tanaka & Farah (1993), which shows examples
of isolated features, intact faces and scrambled faces…………………………………………………….32
Figure 1.8 Example images of one identity with features changed in configuration across
the different images taken from the study by Haig (1984). Features themselves remain
intact, although the exact distances between features are changes……………………………….33
Figure 1.9 Example stimuli from Hole et al. (2002) showing original image, vertical stretch
of 150% and 200%...…………………………..……………………………………………………………..……………34
Figure 1.10 Example stimuli from Collishaw & Hole (2000). Top row, from left to right
images show the following image conditions: intact, inverted, scrambled. Bottom row
images show: blurred, blurred and scrambled, blurred and inverted.……………………….……35

7

Figure 1.11 Example of the card-sorting scenario. Figure shows 20 images of two different
identities. Familiar viewers find this task easy whereas unfamiliar viewers generally
believe that there are more than two identities present. ……………………………………..….……38
Figure 1.12 Example stimuli in the Dhamecha et al. (2014) face-matching task………………44
Figure 2.1 Example of full face and internal feature stimuli (left) and full face and external
features (right) viewed as part of Clutterbuck & Johnston’s (2002) face-matching task….53
Figure 2.2 Face-matching task image examples. The pairs on the left (A) show different
identities (with the imposter face on the right), the pairs on the left (B) show same
identity pairs……………………………………………………………………………………………………………..……57
Figure 2.3 Photograph of one participant’s use of the familiarity scale taken immediately
after completion in the experimental setting. The far left side of the scale indicates that
the face was completely unfamiliar, and the far right depicts extreme familiarity with the
face…………………………………………………………………………………………………………………………………59
Figure 2.4 Percentage of correct responses in face-matching task (using fine quality
200x300 pixel images) for each familiarity quintile; 1 (0-19), 2 (20-39), 3 (40-59), 4 (60-79),
5 (80-100). With Band 1 being completely unfamiliar and Band 5 being extremely familiar.
Error bars show standard error of the mean…………………………………………………………………..60
Figure 2.5 Pairwise comparisons showing which familiarity levels performance was
significantly better than the other familiarity levels..………………………………………………………61
Figure 2.6 Graph showing pattern for Same identity pairs correct response and Different
(lookalike) identity pairs correct responses. Error bars show standard error of the
mean.………………………………………..……………………………………………………………………………………62
Figure 2.7 Example of actual image issued by the police to the public to assist with
identification of a man caught on CCTV (Howarth, 2016). This image takes a pixelated
appearance.……………………………………………….…………………………………………………………………..65
Figure 2.8 Example of stimuli used in the face-matching task created by Bindemann et al.
(2013)..……………………………………………………….………………………………………………………………… 67

8

Figure 2.9 Example of the image appearance for Experiment 2 (top pair) compared with
the fine version of the same image as used in Experiment 1 (bottom) pair. These are
different image pairs of Al Gore with the lookalike appearing on the right..…………………. 69
Figure 2.10 Graph showing the graded effect of familiarity for participants’ face-matching
task performance. Error bars show standard error of the mean……………………………………..71
Figure 2.11 Face-matching performance broken down into correct Same and Different
identity trials. Error bars show standard error of the mean……………….……………………………72
Figure 2.12 Example of coarsely pixelated image stimuli used in Experiment 3………….…..76
Figure 2.13 Percentage of correct responses for each of the three levels of familiarity in
the 20x30 pixel condition. Error bars show standard error of the mean…………………………77
Figure 2.14 Percentage of correct responses in face-matching task (using poor quality
20x30 pixel images) by familiarity broken down into same (dotted line) and different
(dashed line) correct trials. Error bars show standard error of the mean…………………..……78
Figure 3.1 Face shape classification examples provided by Towler et al. (2014)……………. 87
Figure 3.2 Graph from White et al. (2014) passport officer paper showing the officers’
performance accuracy on the GFMT alongside their employment duration. Some police
officers performed very highly on the GFMT, these high scores can be found at both ends
of the employment duration axis……………………………………………………………………………………90
Figure 3.3 Identical images of Al Gore (left) and Gary Barlow (right) shown as they were
presented in each experimental condition. The image on the left for each identity, shows
the coarsely pixelated image as presented in Experiment 3. The images on the right, show
the image on the left of it, after undergoing blurring, and as presented in Experiment 4.93
Figure 3.4 Graph showing performance accuracy on the blurred pixelated task split by
same and different person trials. Error bars show standard error of the mean……………….94
Figure 3.5 Percentage of correct responses in face-matching task for each familiarity band
(low familiarity, mid familiarity, high familiarity), for Experiments 3 (black line) & 4 (blue

9

line). Error bars show standard error of the mean. Error bars show standard error of the
mean………………………………………………………………………………………………………………………………95
Figure 3.6 Mean performance on items of the GFMT performance according to different
crowd sizes (White et al. 2013). Graph shows performance accuracy broken down by trial
type, with results analysed for crowd sizes of 1, 2, 4, 8, 16, 32 and 64..…………………….…100
Figure 3.7 Graph showing the mean accuracy score for crowd sizes of 1, 3, 5 and 15 for
the coarsely pixelated lookalike task, Experiment 3. Error bars show standard error of the
mean…………………………………………………………………………………………………………………………….103
Figure 3.8 Graph showing the mean accuracy score for crowd sizes of 1, 3, 5 and 15 for
blurred version of coarsely pixelated lookalike task (blue line) and the coarsely pixelated
lookalike task (black line).…………………………………………………………………………………..…………106
Figure 3.9 Example trials from the PLT. Images on the left show different identities (with
the imposter face on the right). Images on the right show the same identity............……112
Figure 3.10 Performance of police super-recognisers and comparison viewers.
Performance of super- recognisers (SR1–4; black) and comparison viewers (white) on
three different tests of face recognition—the GFMT (left column), the MFMT (middle
column), and the PLT (right column). Vertical lines indicate the range of scores for
comparison groups, the deleted portion of the line shows the standard deviation, and the
horizontal notch shows the mean. In all three tasks, chance performance is 50%..………113
Figure 4.1 Changes in face shape resulting in differing weight judgments as photographs
were taken from far distance (left) to near distance (right), example taken from Harper &
Latto (2001)………………………………. ………………………………………………………………………………..121
Figure 4.2 Example of measurement figure taken from Burton et al. (2015). Images are
standardised so that interocular distance is the same. Metric distances are expressed as
proportions of standardised interocular distance……………………………………………….….….…122
Figure 4.3 Example of the measurements taken for two of the photos of one model.
Measurements taken were the distances between: left eye to nose, right eye to nose, left
nose to mouth corner, right nose to mouth corner and centre of nose to mouth…..….…126

10

Figure 4.4 Example of one identity with each of their four image identity pairings shown.
The first column shows image pairs of the same identity and the second column shows
different identity pairs. The first row shows same camera-to-subject distance pairs and
the bottom row shows pairs where the images are of different camera-to-subject
distance.…………………………………………………………………………………….…………………………………129
Figure 4.5 Effect of changing camera-to-subject distance on performance accuracy in the
face-matching task for familiar (F) and unfamiliar viewers (U), for same and different
identity trials, at both same and different distances. Error bars show standard error of the
mean………………………………………………………………………………………………………………………….…132
Figure 4.6 Example of congruent and Incongruent face image pairs (with distance cues)
for same and different identities.………….…………………….…………….….…………….……..….….…137
Figure 4.7 Graph showing the percentage of correct responses for both congruent and
incongruent image pairs broken down by same and different identity trials. Error bars
show standard error of the mean………………………………………………………………………………...138
Figure 5.1 Images from the IDV1 database. Props include glasses, fake beards and
moustaches, medical maskes and hats turbans..… …………………………………………………….…148
Figure 5.2 Images from the AR datatbase. Disguise manipulations are limited to a change
of expression or the addition of sunglasses or a scarf..……………….… …………………………….153
Figure 5.3 A sample of images from the National Geographic Database (Ramathan, et al.,
2004) ……………………………………………………………………………………………………………………………154
Figure 5.4 Example images taken from the TarrLab face database. ……………………………..155
Figure 5.5 Examples from the synthetic face disguise database (Singh et al. 2009).……..156
Figure 5.6 Sample of props used to create the disguise face database………………………….159
Figure 5.7 Image taken during stimuli selection process……………………………………………….161

11

Figure 5.8 Example pairs for each condition. Top row shows same identity pairs, the lower
two rows show different identity pairs. Pairs in the first column are in no disguise. Pairs in
the second column are in disguise. All 26 models were photographed in each of the
conditions……………………………………………………………………………………………………………………162
Figure 5.9 Selection of images taken from the disguise base database to create a wheel of
disguise. Images with the same colour frame show the same identity. Images with
different colour frames are of different identities………………………………….…………..….….…163
Figure 5.10 Performance accuracy for unfamiliar viewers for evasion, impersonation
similar and impersonation random pairs when the images consisted of no disguise or
disguise pairs. Error bars show standard error of the mean………………………………………….169
Figure 5.11 Performance accuracy of unfamiliar viewers who were aware of the disguise
component of the face-matching task. Error bars show standard error of the mean…….174
Figure 5.12 Performance accuracy in the face-matching task for viewers who were
familiar with the models whose images featured in the task. Error bars show standard
error of the mean…………………………………………………………………………………………………………180
Figure 5.13 Graph showing performance accuracy for Disguise face pairs for each of the 3
Experiments: U informed (Experiment 10), U uninformed (Experiment 11), Familiar
(Experiment 12) ……………………………………………………………………………………………………………181
Figure 6.1 Figure demonstrates two different disguise images for the same identity. The
image on the left occludes top part of the face, and the image of the right occludes the
bottom part of the face…………………………………………………………………………………………………198
Figure 6.2 Word cloud showing the most frequently stated words for creating an Evasion
disguise…………………………………………………………………………………………………………………………201
Figure 6.3 Word cloud showing the most frequently stated words for creating an
Impersonation Similar disguise. All words represent similarities with the target face
except where specified as differences….…………………………………………………………………….…202

12

Figure 6.4 The model (shown right in this image pair) shaved his beard to better match
the appearance of his target (right)……………………………………………………………………….…..…203
Figure 6.5 The model (right) has copied the eyebrows of the target (left) using makeup to
alter eyebrow shape……………………..……………………………………………………….………..………..…203
Figure 6.6 Word cloud showing the most frequently stated words for creating an
Impersonation Random disguise. All words represent similarities with the target face
except where specified as differences.…….……………………………………………………………………204
Figure 6.7 Graph showing correlation between effectiveness rating and percentage of
errors made for each Evasion disguise item. Data points are spread horizontally if they
would otherwise overlap……………………………………………………………………………………………...204
Figure 6.8 Graph showing correlation between effectiveness rating and percentage of
errors made for each Impersonation Similar disguise Item. Data points are spread
horizontally if they would otherwise overlap…………………………………………………………..……207
Figure 6.9 Graph showing correlation between effectiveness rating and percentage of
errors made for each Impersonation Random disguise item. Data points are spread
horizontally if they would otherwise overlap………………………………………..………………………207
Figure 6.10 Bar graph showing the most frequent forms of disguise for Evasion (blue),
Impersonation Similar (red) and Impersonation Random (grey). Evasion changes capture
differences in appearance with the reference photograph whereas Impersonation
changes represent similarities..………………………………………………………………………………….…210
Figure 6.11 Image example where social inferences were reported to differ between the
reference model image (left) and the model in evasion disguise (right)..……………………..211
Figure 6.12 The model (left) copied the distinguishing feature (mole [on the left side of
the image under the mouth]) of the target (right) by using make up…………………………….212
Figure 6.13 Example illustration of the distance calculations made for the Evasion disguise
condition. Distance moved for incidental change was compared with distance moved for

13

disguise change for each of the 3 disguise conditions (Evasion, Impersonation Similar,
Impersonation Random)……….……………………………………………………………………………………...218
Figure 7.1 Schematic representations of the disguise manipulations with regards to face
space. Each bubble represents one individual’s face space.…………………………..………….….232

14

Acknowledgements

First I would like to express my sincere gratitude to my supervisor Dr Rob Jenkins, for his
continued guidance and support, as well as for his enthusiasm for each and every project
throughout my PhD.

I would also like to thank the members of the University of York FaceLab as well as my
friends and family for all of their encouragement and helpful advice.

Finally, thank you to my disguise models – so much of this thesis would not have been
possible without the phenomenal effort and dedication you put into creating your
disguises. For that I am incredibly grateful. Your involvement helped make this PhD so
enjoyable.

15

Declaration

I declare that this thesis is my own work carried out under normal terms of supervision.
This work has not been previously been presented for an award at this, or any other
University. All quotations in this thesis have been distinguished by quotation marks and
they have been attributed to the original source. All sources are acknowledged as
References.

16

Chapter 1 – General Introduction
1.1 Introduction: Why Face Recognition is Important
Face recognition refers to the ability to correctly verify the identity of a specific individual,
often by comparing a ‘target’ face against other face images. Accurate facial identification
is important because facial image comparison is the basis of many security
infrastructures, such as passport control. Successful face recognition is often critical in the
identification of criminal suspects, and avoiding miscarriage of justice.

The Applied Problem
The need for face recognition research became apparent following several high profile
cases of mistaken identity. In England in the 1890s Adolf Beck was convicted of fraud and
imprisoned as a result of erroneous face recognition. Beck was first imprisoned in 1896,
after being repeatedly picked out in a police line up of face photographs as being the man
responsible for defrauding over 20 women in different attacks. All women had reported
that their attacker was a grey haired man who had a moustache. Beck’s face was the only
face in the line up to have both of these features. After serving his sentence Beck was
released, but soon imprisoned after again being identified (based on eyewitness
testimonies and line up scenarios) for more attacks similar in nature to those prior to his
prison sentence. It was only when these similar attacks continued during Beck’s second
prison sentence that the real culprit ‘Smith’ whose real name was Meyer, was finally
caught. Three of the five women who identified Beck as the culprit at the hearing prior to
his second imprisonment were called in to view photographs of Meyer. All women
admitted their mistake in their previous recognition of Beck and agreed that Meyer was
the true attacker. Meyer had been following the original Beck case and had moved away
to America until Beck served his sentence, returning to the UK and to his previous
fraudulent activity following Beck’s release but was unaware that Beck had been
convicted a second time (Coats, 1999).

17

Similar high profile situations of mistaken identity still exist in the United Kingdom. On the
22nd July 2005, a case of mistaken identity cost Jean Charles de Menezes his life. The
metropolitan police mistook Jean Charles de Menezes, a Brazilian electrician, for Hussain
Osman who was a suspect involved in failed bomb attacks, which had been carried out in
London the previous day. Several incidents led to the police shooting Jean Charles de
Menezes eight times, believing that they were faced with Hussain Osman at the time of
shooting. One of the contributors was erroneous face-matching - police had been given a
photograph of Osman, and mistook Jean Charles de Menezes to be him (see Figure 1 for
example images of the suspect and mistaken suspect) (BBC News, 2005; Cowan, 2005).
This tragedy echoed the message that when face recognition goes wrong, very serious
consequences can occur.

Figure 1.1 Image showing the face of the real suspect, Hussain Oman on the left and the face of Jean
Charles de Menezes on the right.

In addition to these famous cases, several studies highlight the weight that eyewitness
face recognition can hold on a jury verdict. Wells et al. (1998) found that out of the first
40 cases where later DNA evidence excluded wrongfully convicted suspects as the culprit
of a crime, 90% of these wrongful convictions had been strongly based on evidence from
erroneous eyewitness testimonies. Additionally, Huff (1987) found that mistaken
eyewitness identification of a convicted culprit occurred in 500 wrongful convictions that
were later investigated.

18

Loftus & Doyle (1992) tested the weight given to eyewitness testimonies experimentally
by creating a mock trial situation. Two separate sets of jurors took part, each set were
presented with the same evidence, however one of the sets of jurors were also given
evidence from an eyewitness testimony whereas the other set of jurors received no
eyewitness testimony. When no eyewitness testimony was presented, 18% of jurors gave
a guilty verdict. This jumped to a guilty verdict of 72% for those who received the
eyewitness testimony. Further still, even when the eyewitness testimony was said to be
inaccurate, 68% of jurors delivered a guilty verdict. Jurors tend to overestimate the
accuracy of eyewitness testimonies (Brigham & Bothwell, 1983). It is important to
understand whether these cases of mistaken identity are unfortunate anomalies in face
recognition or whether they are indicative of a very serious problem – people may not be
as good as they think they are at recognising faces. A great deal of research has been
done on this, which will be discussed later in this chapter.

Solutions to the Face Recognition Problem
The examples and studies discussed above clearly demonstrate that there are situations
where face recognition has been inaccurate in the past. Several procedures and
technologies exist to aid identification. For example, passports, driving licences and
university student cards feature an image of the holders face. This is intended to allow a
viewer to make an identity judgment by comparing the image of the physically present
cardholder with the image on the card that they are holding. If the images are considered
a match, then the holder will be granted the access or purchase that they wish to acquire.
If the images are not considered to match this would imply that the holder is using
fraudulent identification, and the carrier would not be granted their access or purchase
request.

Additionally, police line-up situations exist to try and reduce the number of erroneous
face identifications in an investigation. In a line-up situation, witnesses of a crime view

19

the faces of several people, often of a similar appearance, one of whom is the suspect in a
criminal case. The witness is tasked with picking out the suspect from the line-up array of
faces. Line-up situation errors existed in several of the cases discussed above,
demonstrating that this method of facial identification does not always render accurate
results.

Closed circuit television (CCTV) cameras have become more prevalent in recent years
meaning that more facial images are stored now than ever before. CCTV camera footage
is used in many forensic investigations as the images can both capture the image of a
suspect and track the suspect’s movement. CCTV camera footage is however not always
of good quality, and images available for comparison may come from multiple cameras,
with breaks in the footage where CCTV is not in operation.

Demands Experienced by Human Viewers
Although there are several methods in place, which aim to ensure accurate identification
of an individual occurs, each of the methods places demand on the human viewer who is
required to make the identity judgment from the visual information available. Document
holder to card image identity comparisons and CCTV camera footage evaluation require a
viewer to make an identity judgment by comparing the images presented before them to
determine whether the identities are the same person or different people across the
multiple images. Line-up scenarios include a memory component – a witness has to
compare the previously seen image of a face, which they have stored in memory, with the
images seen before them. This memory component to the task may affect face
recognition accuracy. Human ability to meet each of these demands is assessed below.

Human Face Memory Ability
The famous cases of mistaken identity discussed earlier involved aspects of erroneous
face memory. To test face memory performance experimentally, Bruce (1982)

20

investigated human face memory performance for identical images, images changed in
either pose or expression and images changed in both pose and expression. Participants
in the study viewed images of 24 people, and were told to try and remember the face of
the person in each of the images viewed as they would be asked to identify the images in
a later recognition task, in which the identity may be shown in a changed image.
Recognition accuracy was highest for the unchanged images (90% rate), followed by the
images changed in pose or expression (76% hit rate) and finally images changed in both
pose and expression (60.5% hit rate). It is important to note that in real world face
recognition tasks the images for comparison are always changed between learning and
viewing the face at a later stage. This study provides experimental evidence for poor face
memory performance for changed face images. Alongside the example of famous cases of
mistaken identity discussed above, it is clear that face recognition involving face memory
is problematic.

Perceptual Matching
Initially, memory was thought to be the whole problem of face recognition. However, in
recent years, identification by face-matching has become important. In face-matching the
task is to compare two images—both of which are physically present – in order to make
identity judgments. This is the scenario used in passport control, other identity checks
involving photographic identification, and also by police to piece together an incident
through comparison of multiple CCTV images. There is no memory component in this task
as all of the information needed to make a decision is available in the images presented.
However, it soon became clear that the problems of face recognition surpassed
erroneous memory.

1.2 Familiarity & Face-Matching
Face-matching accuracy differs greatly for faces that a viewer is familiar with compared to
performance for faces that are unfamiliar to the viewer (e.g Burton, Wilson, Cowan &
Bruce, 1999). This section will examine face-matching performance for both unfamiliar

21

viewers and familiar viewers in turn and then explore the idea that familiarity is more
than a binary variable; investigating research which treats familiarity as a graded concept
(Clutterbuck & Johnston, 2002,2004,2005).

Unfamiliar Face-Matching
People are extremely accurate at matching two identical images of a face (Bruce, 1982).
Face recognition, as required by the security and forensic situations outlined above, relies
on the ability to match or identify a mismatch between different images of faces. The
literature covers several unfamiliar face-matching scenarios including person to
photograph matching (Kemp, Towell & Pike, 1997; White, Kemp, Jenkins, Matheson &
Burton, 2014), police line up scenarios (Burton, Wilson, Cowan, & Bruce, 1999) and paired
image comparisons (Megreya & Burton, 2006, 2007; Burton, 2010).

Unfamiliar Live Face-Matching
The comparison of a face photograph to a physically present face is the method of
identity verification that matches that of passport control, building access, shop sales and
many other security situations. There is no memory component in such a task as both the
physically present cardholder and the identification card are available for direct
comparison. Kemp et al. (1997) tested performance accuracy for matching an ID card
holder (physically present) to a photo on the ID card. The experiment took part in a large
supermarket in the outskirts of London where participants in the study were experienced
supermarket cashiers. The cashiers’ task was to determine whether the photograph on
the identification cards portrayed the holders. Each cashier judged physical appearance
against the identity cards for an average of 44 shoppers with whom they interacted at the
checkout (interaction was as in a normal shopping scenario). Identification performance
was poor - the cashiers performed with 67% accuracy. More than half of fraudulent cards
were accepted as the holder’s true identity. In this experiment identification card
photographs were taken just six weeks before the experiment. In the UK, passport
photographs are valid for up to ten years before the photograph needs updating. This

22

means that the cashier task could be even easier than real life situations of cardholder to
image identity matching as a face may look even more different over longer periods of
time.

While performance was poor for the supermarket cashiers, one might expect trained
personnel who work in security to excel at a live person to image matching task. White,
Kemp, Jenkins, Matheson & Burton (2014) recently tested passport officers face
recognition ability in a task similar to Kemp and colleagues (1997) study. Passport officers
were tested on their ability to match live subjects to a passport style photograph that was
either of the same identity or showed an image of a different identity. Just 16 female, and
16 male student volunteers took part as the photo holders and identities for this
experiment. Different person trials were created though pairing together the most similar
images from the images provided by the volunteers. As the volunteer pool was small, this
greatly limited the ability to provide an extremely convincing foil. Additionally, all
photographs were taken just two weeks before the experiment was carried out. These
conditions may make the experiment easier than a real life situation would be, as in real
situations of identity fraud people may have a large sample of available different person
passports to chose from, and additionally for same person trials, people may look more
different in their real passport photograph than they did in the image used in this task as
less time had passed since the time of photograph, leaving less time for natural changes
as a result of aging. Nevertheless, performance in the matching task was poor. On
average, passport officers made errors in 10% of cases, with 14% of different identity
pairs being mistakenly accepted as same people. Years of experience and training in
matching faces made no difference to performance. Despite overall levels of poor
performance, there were some individuals who performed very well on the task (see
Figure 1.2). A photo-to-photo matching test was then conducted to compare passport
officer performance with that of undergraduate student participants. Student’s
performance was indistinguishable from that of the passport officers. This study
highlights the difficulties of live-person to image matching, and also demonstrates that
passport officers, a highly trained group, are no better than undergraduates at facematching.

23

Figure 1.2 Face-matching performance of passport officers in the study by White et al. (2014). Some officers
perform with very high performance regardless of their employment duration.

Line-ups
In addition to live person to image matching, poor face-matching performance has also
been reported in experiments that replicate the police line up scenario. Bruce,
Henderson, Greenwood & Hancock (1999) created a face-matching task based on a police
line up situation. Participants had to match a given photograph of a face to one of the
face images from a line up of 10 face images presented in an array below the target face
(see Figure 1.3). Participants were told that the target face may or may not be present in
the array line-up, as would be the case in a real criminal line-up situation. Notably, unlike
a real line-up situation, no memory component was involved for the viewer of the faces.
All face photographs were taken on the same day, in the same lighting, and in the same
pose. The only deliberate modification between images was that the target photograph
was taken with a different camera to the array photographs. Despite the consistencies
between the target photograph and the correct match in the array line-up, performance

24

was poor. For both the target-present and target-absent array scenarios, accuracy was
70%. Performance accuracy for correctly matching a target to the same identity present
in the array remained poor even when participants knew that the target face was present
in the array (Bruce et al. 1999; Bruce, Henderson, Newman & Burton, 2001).

Figure 1.3 Example of line stimuli taken from Bruce et al. (1999) line up task. The correct match of the
target face is face number 3.

Paired Face-Matching
More recently, work on face-matching ability has focused on paired matching paradigms,
in which viewers make identity judgments (same or different) for two faces presented
simultaneously. The viewer’s task is to say whether the two faces in the pair presented
depict the same person or different people. Face-matching by image is a good model of
real world tasks and has important practical applications. For example, police
investigations often require accurate matching of an image taken from CCTV footage to
another image of the suspect, such as a mug shot. Alternatively, two different face images
taken from different CCTV footage may be compared in order to track a person’s

25

movement. As in the experiments described above, the face-matching paradigm involves
no memory component – both images are viewed simultaneously and the viewer’s task is
to say whether the faces presented are of the same person’s face or are faces of two
different people.

To provide a standardised measure of face-matching ability, Burton, White & McNeil
(2010) designed a task known as the Glasgow Face-Matching Test (GFMT). The images
used to construct this test were taken at a constant viewing distance and same camera
angle, under constant lighting conditions, and the expression of the face was unchanged
across photographs of the same identity (see Figure 1.4 for examples). Photographs of the
same individual were taken only minutes apart, with no deliberate attempt to change the
model’s appearance in any way between photographs. Thus, images should have been of
optimal condition to allow accurate comparison. Different person trials were created by
pairing the faces with the most similar other person’s face image out of those available.
Two versions of the GFMT were created – the full version (160 items), and a short version
consisting of the 40 most difficult items from the full version. Participants made over 10%
of errors in the full version. This performance is considered low, especially considering the
controls taken to make the images as similar to each other as possible and given that
participants experienced no time limit while completing the task. The short version of the
task rendered poorer results, with participants making nearly 20% errors.

26

Figure 1.4 Examples of face pairs taken from the GFMT (Burton, et al. 2010) Top row show different identity
pairs, bottom row are same identity pairs.

In summary, these studies successfully and repeatedly demonstrated the real world
problem of poor face-matching in a lab environment. However, these findings go against
most people’s intuition. People tend to believe that they are good at face recognition
(Jenkins & Burton, 2011). It is important to remember the vast majority of face-matching
that people are faced with on a daily basis involves familiar faces. People experience
regular and repeated success in recognising family members and friends, and do so with
very little difficulty.

Familiar Face Matching
People are extremely good at matching familiar faces, even when image quality is poor.
To test this, Burton, Wilson, Cowan & Bruce (1999) designed a face-matching task using
poor quality CCTV footage (see Figure 1.5). This matching task was presented to people
who were familiar with the faces in the CCTV footage, and also to a group of unfamiliar
viewers. Familiar viewers could match the faces in the task with almost perfect levels of
accuracy, whereas unfamiliar viewers performed at chance level or highly inaccurately on
the exact same task (Bruce, Henderson, Newman & Burton, (2001); Burton, Wilson,
Cowan & Bruce (1999).

27

Figure 1.5 Examples of face pairs in the matching task involving poor quality CCTV footage in the
experiments conducted by Burton et al. (1999). Familiar viewers performed with high accuracy when
matching the face pairs.

This familiarity advantage can be achieved quickly. Megreya and Burton (2006)
familiarised participants with faces by showing participants 30 second long video clips of
the 40 identities that the participants were told to try and learn as they would see the
faces again in a subsequent face-matching task. Even after this brief familiarisation with a
target face, participants performed better in the task in trials involving familiarised faces
over novel faces. However this advantage only existed for upright faces. When the
familiarised face images were inverted, participants performed better on the unfamiliar
upright faces than the inverted versions of the familiarised face images. This is likely due
to the range of experience that a viewer has with a face and is a topic which will be
addressed later in this chapter. Megreya & Burton (2006) concluded that unfamiliar faces
are “processed for identity in a qualitatively different way than are familiar faces”. The
authors suggested that unfamiliar faces are processed in a similar way to other objects
whereas we have a special ability for faces when they become familiar, which relies on a
different processing system.

28

Familiarity as a Graded Concept
Many past studies have classified faces as either familiar or unfamiliar. Recently,
experiments have demonstrated a graded nature to familiarity, with familiarity with a
face increasing as exposure to that face increases. To test for a graded nature of
familiarity, Clutterbuck & Johnston (2002) divided the familiarity of faces viewed in their
study into 3 levels, rather than treating familiarity as a binary variable. Faces in
Clutterbuck & Johnston’s (2002) study were described as being highly familiar, of medium
familiarity and unfamiliar. Familiarity distinctions for these celebrity face images were
made by a group of eight independent raters. The high and medium familiarity group
were created based on the mean familiarity ratings for each face. Unfamiliar faces were
faces that were not famous and were not known by the participants. The results
demonstrated that participants were fastest at accurately matching full faces to internal
features of a face for the faces for the highly familiar faces in the study (Figure 1.6). Speed
of correct response decreased as familiarity level decreased across each of the three
familiarity groups. All of the faces used for the familiar and medium familiarity groups
belonged to celebrities. By breaking familiarity into 3 levels this study begins to separate
the level of familiarity of these celebrities which begins to address the issue that not all
celebrities are equally well known. However, this method fails to address that different
participants in the study may be more or less familiar with different faces to each other.

Figure 1.6 Examples of full-face image on the left was to be matched with either of the internal face images
on the right. Familiarity aided matching accuracy on this task.

29

Clutterbuck & Johnston (2004) also demonstrate a graded familiarity advantage for
accuracy in a gender judgment task. The authors again distinguished between three levels
of familiarity but constructed the familiarity groups in a different way to their previous
(Clutterbuck & Johnston, 2002) study. Faces in the gender judgment tasks were either
previously familiar (know to the participant before the experiment), of acquired
familiarity, or unfamiliar (never seen before the testing phase of the experiment).
Participants achieved acquired familiarity by viewing sets of previously unfamiliar images
for two second periods, with each face image viewed a total of ten times. The results of
the study were that participants were best at the identity judgment task for faces which
they were previously familiar with (familiarity was measured by eight independent raters
as in Clutterbuck & Johnston, 2002), and gender judgement performance was of higher
accuracy for acquired familiarity faces than for the unfamiliar faces. This acquired
familiarity method gave a similar pattern of results for face-matching speed (Clutterbuck
& Johnston, 2005). Participants were fastest at correctly rejecting mismatch faces when
they were previously familiar with them, followed by those for which they had learnt as
part of the experiment, and slowest for faces which were unfamiliar at test. The three
experiments conducted by Clutterbuck & Johnston (2002, 2004, 2005) suggest that face
recognition is influenced not only by whether a face is familiar or not, but by the level of
the viewer’s familiarity with the face.

1.3 What Information is used to Recognise a Face?
The literature has shown that people get better at recognising faces when they learn the
faces in question, but it is unclear what information from an image viewers actually use to
make identity judgments of faces.

There has been much debate in the face recognition literature over which information
from a face allows the face to be recognised. The main lines of debate over the
information used within a face to aid identification, focus on a featural versus configural
processing account. Featural processing involves scanning the face for features (e.g. eyes,

30

nose, mouth), and then using the information from the detail of these features to decide
whether the face is known. Configural processing on the other hand, argues that the
spatial layout between features is what is important for successful face recognition.
According to the holistic processing theory, faces are identified ‘as a whole entity’, with
both a match in configural and featural information being important for recognition
(Sergent, 1984; Bruce, Hellawell & Hay, 1987). Studies have generally taken the approach
that if a manipulation applied to a face image slows down and/or impairs face
recognition, then whatever was changed about the face was important for the face
recognition process (Donders, 1868/1969).

Featural Accounts
Bradshaw & Wallace (1971) provided one of the first studies to suggest that face
recognition is based primarily on the recognition of facial features. Their experiment
comprised of images of identikit faces - faces made up of face elements [individual
features], which can be combined to make a face. This method is used by police to create
an image of a suspect based on a witnesses’ description of the suspect’s face. Participants
in Bradshaw and Wallace’s (1971) study viewed pairs of identikit faces and judged
whether the image pairs presented showed the same face or different faces. Different
parings differed in the number of shared features in the two images. For different person
trials, accurate identity judgments were made more quickly as the number of differences
in features between the faces increased. On the basis of this result it seemed that
participants scanned the face images until they found a mismatch of features across the
image pairs. Thus, it was concluded that the results were best explained through a
featural account of face recognition. An earlier experiment by Smith & Neilsen (1970)
supported this result. They reported that as the number of featural differences between
pairs of faces was increased, there was a decrease in the time needed to tell the two
faces apart. This suggests that participants were using a feature comparison technique to
aid their identity judgment of the face images. These studies looked only at performance
for unfamiliar faces, therefore they do not provide information on what is learnt during
the familiarisation process with a face.

31

Figure 1.7 Example of test stimuli used by Tanaka & Farah (1993), which shows examples of isolated
features, intact faces and scrambled faces.

Further evidence for a featural processing account comes from studies involving
scrambled faces. Scrambled face images contain all the featural information as in an
intact face image, however the features appear in a jumbled order. Tanaka & Farah
(1993) tested whether participants were able to recognise learned identities from
scrambled face images, or even from viewing a feature from a face in isolation of the
whole face. They reported that participants made identity judgments with moderate
accuracy in the context of a scrambled face and also when features belonging to a
previously seen identity were viewed in isolation of the rest of the face (see Figure 1.7).
This suggests that features, regardless of their configuration or even the presence of the
rest of the face, provide information which aids face recognition (Bruyer & Coget, 1987;
Tanaka & Farah, 1993).

Configural Processing
In addition to the research that argues for the importance of featural information in face
recognition, there is a wealth of research that configural processing is important.
Theories of configural processing suggest that the layout of features within a face (1st
order configural processing), specifically the metric distances between features in a face
(2nd order configural processing) are learnt and used in face recognition. For example,
Haig (1984) tested the effect of digitally editing the distance between features of a face,
without editing the features themselves in any way, on face recognition performance.
Haig (1984) found that even when featural information was present in a highly accessible

32

form (see Figure 1.8 for example stimuli), recognition for unfamiliar faces was severely
dampened when subtle changes were applied to the relationship of the layout of the
facial features. This supports theories of configural processing.

Figure 1.8 Example images of one identity with features changed in configuration across the different
images taken from the study by Haig (1984). Features themselves remain intact, although the exact
distances between features are changes.

Many studies on this topic are based on a similar logic; any manipulation that is seen to
reduce recognition accuracy or speed is thought to be important for the recognition
process. It is believed that configural information is harder to access in an upside down
image than in the upright form that people have everyday experience with (Yin, 1969).
Familiar orientation had been found to be important for past studies involving object and
letter recognition (Henle, 1942; Ghent, 1960), and it is argued that the familiarity factor
for upright orientation also exists for face recognition (Yin, 1969). Therefore, if configural
information is important for face recognition, performance accuracy would drop in an
inverted face memory task, as configural information is more accessible in an inverted
than upright face image. Yin (1969) found exactly that result. In general, people show a
bias for remembering upright objects, with various upside-down objects proving more
difficult to remember (Goldstein, 1965; Hochberg & Galper, 1967), however, faces were

33

disproportionally affected by inversion, with poorer results for remembering faces than
objects in the inverted image memory task (Maurer, Grand & Mondloch, 2002; Rossion,
2009). Poorer recognition performance for inverted than upright faces suggests a role for
configural information in face recognition. Again, this finding was for unfamiliar faces.

Despite the arguments for configural processing provided by the studies discussed, the
usefulness of the content has been queried. Several studies show that configural
information in a face can be changed, and faces still successfully recognised. Hole,
George, Eaves & Rasek (2002) investigated whether familiar face recognition was affected
by drastic changes in the featural layout of the face. They reported that faces could
undergo vertical stretching of twice their original height, with no change in face
recognition performance for these face images (see Figure 1.9 for example of image
stretching). Bindemann, Burton, Leuthold & Schweinberger (2008) added to this finding,
as they demonstrated that the N250r ERP response to faces is unaffected by stretching of
the face. These findings suggests that neural processes involved in face recognition, and
also behavioural face recognition, are unaffected by the configural changes caused by
stretching, suggesting that consistent configural information is not important for these
processes.

Figure 1.9 Example stimuli from Hole et al. (2002) showing original image, vertical stretch of 150% and
200%.

The configural account has also rendered failure in practical application. Configural
processing has failed to provide adequate facial identification results in both early

34

automated face recognition systems (Kelly, 1970; Kanade, 1973), and more recently
Kleinberg, Vanezis, & Burton, (2007) presented evidence that anthropometry methods fail
to identify targets in forensic investigations.

Are Both Important?
Collishaw & Hole (2000) investigated the effect of a range of face image manipulations on
face recognition performance for both familiar and unfamiliar faces (Figure 1.10). Two of
these manipulations – scrambled and inverted (upside down) faces – disrupted the
configural information in an image, but kept the features themselves intact. Another –
blurring – made it very difficult to access the featural information in a face. It was found
that the identities in the image, regardless of the viewers familiarity with the identity,
could be recognised ‘much of the time’ in all of these situations. Thus, suggesting that as
long as one of these processing methods is available, face recognition can take place.
Further image conditions – blurred and scrambled, blurred and inverted – were
recognised at levels of around chance. This finding reaffirms the need for one of these
routes of recognition to be present for successful recognition to occur.

Figure 1.10 Example stimuli from Collishaw & Hole (2000). Top row, from left to right images show the
following image conditions: intact, inverted, scrambled. Bottom row images show: blurred, blurred and
scrambled, blurred and inverted.

35

In summary, evidence points to neither featural or configural accounts providing a full
answer to the face recognition problem for both unfamiliar and familiar face recognition.
Images can be difficult to recognise when featural information is present but
configuration is changed, and recognition performance can also be poor when configural
information is intact, such as in photo-negatives and sketch drawings (Galper, 1970;
Davies, Ellis & Shepherd, 1978; Rhodes, Brennan & Carey, 1987; Bruce, 1992; Kemp,
1996). It is also possible that recognition relies somewhat on texture information –
something ignored by the configural and featural accounts (Kemp, 1996; Bruce, 1992).

1.4 Learning Variability
It is possible that what is being learnt when becoming familiar with a face, is not in fact a
specific factor such as the exact details of features themselves or of facial configuration,
but instead an accepted range of appearances for any given face. People are very good at
recognising face images for people they are familiar with, providing any new image is
similar to those previously experienced, but are poor at recognising people in situations
beyond this experience (Bruce, 1994). People vary in appearance naturally across multiple
images. This can be a result of changes in expression, pose, hairstyle, skin complexion,
clothing and many other variables. Bruce (1982) demonstrated changes in pose or
expression between items learned and presented at a later recognition test, lowered
recognition accuracy for unfamiliar viewers. Unfamiliar viewers could rely only on the
information from their one previous exposure with the face image, and could not
generalise this exposure to changes of expression and viewpoint. Familiar viewers
performed accurately on the task despite these image changes, as they could call on a
large number of previous encounters of the face under different conditions to aid their
identity judgments (Bruce, 1982; Burton et al. 1999). Familiar viewers cannot however
generalise past experience with a face to aid recognition in the case of novel face
transformations. For example, familiar viewers have been found to perform poorly in
recognition tasks of photo negative faces and images changed in pigmentation (Bruce &
Langton, 1994) – these changes fell outwith the previous range of experience acquired for
the familiar faces.

36

Further evidence to suggest that familiar viewers store an accepted range of experiences
from a face come from studies on facial configuration. Bruce, Doyle, Dench & Burton
(1991) presented participants with a series of face images which were generated using
facial image software Mac-a-Mug. Different versions of the same face image were created
to provide various configurations of the same face (e.g. the features could be moved up,
down, further apart or closer together). Participants were later presented with a series of
test image pairs, and asked to identify which image from each of the pairs was identical to
an image that they had seen earlier. Accuracy was at ceiling levels for recognition of an
identical face. Faces where configurations lay outwith that in the previously seen face
images were not recognised at test. However, the central image of those previously
viewed (although never itself presented), was also rated as highly familiar, demonstrating
that familiar viewers would accept face images which fell inside the range of face images
they had experienced for a face, but reject any faces which fell outside this experienced
range (Bruce et al., 1991). More recently, Sandford and Burton (2014) asked participants
to resize an altered image of a face to make the face ‘look right’. Familiar viewers were no
more accurate than non-familiar viewers at this task. If configuration does not remain
constant for faces across various images, then it makes sense that there would not be just
one accurate configuration, but a range of accepted configurations that could be stored.

The effect that learning variability has on face-matching accuracy is demonstrated in a
study by Jenkins White, Montfort & Burton (2011). First they highlighted the difficulty
that variability between photographs of the same face causes in unfamiliar facematching. Jenkins and colleagues (2011) challenged participants to separate images of
cards into piles according to identity. Images of the same identity were to be placed in the
same pile. There were in fact only 2 identities - 20 images of each. The 2 identities were
Dutch celebrities who were well known in the Netherlands but unfamiliar faces for British
viewers. Unfamiliar (British) participants averaged rating of the number of identities
present among the cards was 7.5. Familiar viewers (Dutch participants) were then asked
to do the exact same task, and easily and accurately separated the faces into the two
identities present. This study highlights the range of variation that familiar viewers have

37

stored for an identity and the advantage this holds in identity matching. Figure (1.11)
demonstrates some of the many ways that the same faces can vary across images.

Figure 1.11 Example of the card-sorting scenario. Figure shows 20 images of two different identities.
Familiar viewers find this task easy whereas unfamiliar viewers generally believe that there are more than
two identities present.

Face Space Theory
One influential concept for understanding face recognition by variation is the theory of
face space. It is possible that when a new face is encountered, or a previously viewed face
takes on an appearance different to the way it looked before, the mental representation
for that face image is then stored in that person’s ‘face space’. Valentine (1991) was first
to propose the face space model for faces. Valentine argued for a single, although
multidimensional, face space, in which faces for all identities are stored. Faces could be
stored as a single averaged image for each identity, or individual face images could all be
stored within this face space. Valentine believed that faces were organised within this
space by dimensions specific to faces. Specifically, faces were proposed to cluster within
face space in ways which explain phenomena of face recognition including the own race
bias, and typicality effects (Valentine, 1991). There are however shortcomings of this
proposal, for example Burton & Vokey (1998) argue that most of the faces within

38

Valentine’s proposed face space are located away from the centre of face space,
therefore there are very few ‘typical’ faces.

An alternative formulation of face space would be that instead of one face space
encapsulating all faces, in order to account for variation within each individual face, each
face may have its own face space. Each face space could be thought of as a
multidimensional ‘bubble’ that encapsulates all the ways and range of ways that one face
is accepted to look. Any new image would then be compared against the representation
held in face space, for that face, and against all other stored identities’ face spaces, in
order to make an identity judgment. Expertise would be acquired one face at a time, not
for all faces as a class of stimuli.

This formulation is in line with the familiarity advantage, as familiar viewers have had
more encounters with a face or been exposed to the face across more varied viewing
conditions than an unfamiliar viewer. Thus, when a person is becoming familiar with a
face they are learning the different appearances that the specific face can take, and this
refines the face space held for that individual. If, as I suggest, identity judgments are
made by comparing a new image with the representation of a face stored in that person’s
face space, then identification accuracy should increase in alignment with refinement of
face space, i.e. familiar viewers will be more accurate as they have a more refined face
space than unfamiliar viewers. This means that false match (different identity) items are
more likely to fall outwith the range of accepted faces for the target face for familiar than
unfamiliar viewers, and true (same identity) images of target will be more likely to be
accepted to fall within the accepted range for the target face. It appears that exposure to
greater variation with a face increases the likelihood of later correct identity judgments
being made (Menon, White & Kemp, 2015).

Within person variability is a strong component of this face space model, as an identity
specific face space would encompass the idea that people’s appearance naturally changes

39

across images. Sometimes people can look unlike themselves, or similar to another
person, just by chance. This incidental variation between faces is very different to
deliberate change in appearance. Whilst people can look incidentally unlike themselves,
or like another person they can also make a deliberate attempt to look unlike themselves
(evasion disguise) or like another person (impersonation disguise). These appearance
changes, both incidental and deliberate (disguise) will be a key theme of challenging
images explored in this thesis.

1.5 Face Recognition in Challenging Situations
Past research has demonstrated that people are very good at matching two identical
images of a face and that problems instead arise when dealing with matching multiple
images of the same face (Bruce, 1982). When the target faces are unfamiliar to viewers,
performance on such tasks is very poor (Kemp et al. 1997; Burton et al. 1999; Megreya &
Burton, 2006,2007; White et al. 2014). However familiarity with the target faces aids face
recognition accuracy in a range of situations (Burton et al. 1999, Bruce et al. 2001,
Megreya & Burton, 2006, Jenkins & Burton, 2011). The studies discussed up until this
point have tested face recognition where the people who are photographed to provide
images for identity comparisons cooperate with the identity effort, i.e. for same person
trials the models made no deliberate attempt to change appearance across the different
images taken. Furthermore, different person trials were created by selecting the most
similar face images from a small number of available images, therefore the different
identity images may not look all that similar to each other. Face-matching performance
could be even worse for more challenging identity matching scenarios.

As demonstrated above, people can look very different to themselves across different
images (Jenkins, 2011). There are also situations where people naturally look very similar
in appearance to another person. In addition to incidental change, in real world scenarios
people may have strong incentives to create deliberate changes to their appearance -
either to evade their own identity (evasion) or to impersonate someone else
(impersonation). There may be reasons for a person to hide their identity (evasion) for
example if they had been banned from a place but wanted to gain access. Impersonation

40

disguise is also a serious issue. Criminal activities including illegal immigration, the
smuggling of drugs, weapons or stolen goods, human trafficking and terrorist activity may
involve the use of stolen identity documentation in order to cross borders as someone
else. People might choose a passport based on natural similarities with the face. It is also
possible that the new document holder will make an effort to look like the photographed
face on the document in order to reduce suspicion and successfully pass borders using
the stolen identity. Such attempts are made quite frequently, for example, in 2010, 359
people were found guilty in the UK for possessing false or improperly obtained ID. Under
the 2006 Fraud Act an additional 7,319 were charged for dishonestly making a false
representation to make gain for oneself or another (Home Office, 2012). There have been
some attempts to investigate the effect that i) natural/incidental image manipulations;
and ii) deliberate disguise manipulations, have on face recognition performance. These
will be discussed below.

Natural Image Manipulations
It is likely that there are certain image manipulations that increase the likelihood of these
challenging image situations arising. Manipulations including changes in pose and
expression have already been found to make identity judgments more difficult across
images (Bruce et al. 1982).

It is possible that another simple image manipulation, changing camera-to-subject
distance between comparison images, could make same person identity judgments more
difficult than comparing images taken at the same distance. Harper & Latto (2007)
showed that changing camera-to-subject distance changed perceived weight judgments
of a face. This manipulation has not yet been experimentally explored with reference to
face recognition ability, however the results suggest that images changed in camera-tosubject distance would be more difficult to ‘tell together’ as perceptions of the face
changed as a result of camera-to-subject distance change.

41

Deliberate Disguise
Most of the research on recognising disguised faces has focused on changes to featural
information in the face, with reference to evasion disguise. Patterson & Baddeley (1977)
investigated whether face recognition was affected by disguise presence. In this study
participants viewed (with the intent of learning) a series of face images. Participants were
later shown a series of test images, some of which showed the previously seen identities
in disguise, and were tasked with recognising previously seen identities from these
images. Disguise manipulations reduced recognition accuracy to approximately chance.
This task was an image memory task, where the exact images were used at test and
learning, except for those images in the disguise condition. Disguised images included the
addition of props to a face, such as wigs, glasses, fake moustaches and beards.

Terry (1993) found that a more specific and simple form of disguise, the addition of
glasses to a face, was in itself detrimental to face recognition performance. However, this
study reports an effect of eyeglasses on face image memorability, rather than an
experimental disguise manipulation. Participants viewed images of faces (some with
glasses and some without) and were later asked to identify the previously seen images
amongst distractor images. There was no addition or removal of glasses between images.
This was an image memory task rather than a face recognition experiment. In a later
experiment Terry (1994) approached the eyeglasses manipulation from a more controlled
angle. This time, participants learnt face images and were then tested on face images that
could have had glasses or beards added or removed to the image. The removal of glasses
on a person who had initially been presented wearing glasses, and the addition of a beard
lowered recognition accuracy (Terry, 1994).

Furthermore, Righi, Peissig & Tarr (2002) suggested that some disguise manipulations
were more detrimental to memory performance than others. They reported that
recognition performance was hindered if the face image changed in any way between the
learning and test phase. However, recognition was significantly worse when the disguise

42

manipulation involved a change of hairstyle or removal of glasses, compared to when just
glasses were added to a face.

The effectiveness of different types of disguise may depend on the viewer’s familiarity
with the disguised face. External features of a face have been found to be of particular
importance in unfamiliar face recognition (Bruce et al. 1999; Bonner, Burton & Bruce,
2003; Megreya & Bindemann, 2009), whereas internal features may be of greater use in
familiar face recognition (Ellis, Shepherd & Davies, 1979; Young, Hay, McWeeny, Flude, &
Ellis, 1985). To date, disguise studies have not looked at the internal/external feature
manipulation in terms of effectiveness of disguise type and familiarity with the disguised
face.

The studies of disguise discussed so far have all involved a face memory component.
Dhamecha, Singh, Vatsa & Kumar (2014) provide the only published study to date to
investigate people’s ability to face match disguised faces. Photographic models in the
study were given a range of accessories which they could use as they wished to disguise
themselves. Participants then completed a printed questionnaire, which showed pairs of
face images, and were tasked with deciding whether the images in the pair were of the
same person or not. Participants viewed faces of both the same and different ethnicity to
themselves, and also familiar and unfamiliar faces. Highest accuracy rates were found for
same ethnicity, familiar faces. This study provides a very interesting first look at facematching ability for disguised faces, in particular it is interesting that familiarity aided
recognition performance. Although this study provides a recent focus on the problem of
face-matching with disguised face, the stimuli in the study may not be realistic (the sorts
of manipulations that a person would naturally chose to disguise themselves) and
certainly not undetectable disguises (see Figure 1.12). Many of the disguise manipulations
occluded features of the face. Occlusion disguise would not be effective in evading
identity in all situations, for example passport security checks often request that items of
occlusion are removed from a face during an identity check. Furthermore, this study
focuses exclusively on evasion disguise. This is a very interesting question to address as

43

people may have very strong reason to not be identified as their true self, however
evasion disguise only covers half of the disguise problem. If a person is travelling on a
stolen passport, it is more likely that they would attempt to make themselves look
specifically like the person on the passport rather than only trying to hide their own
identity. Given the threat of successful disguise to security it would be useful for studies
to investigate realistic disguise for both evasion and impersonation situations.

Figure 1.12 Example stimuli in the Dhamecha et al. (2014) face-matching task.

In summary, where disguised face recognition performance has been investigated, focus
has been exclusively on evasion disguise (when a person changes their own appearance
to look unlike themselves). Stimuli used in these studies have focused on simple disguise
manipulations, mostly the addition of props to occlude parts of the face. Furthermore,
experiments have compared disguise face recognition performance with exact image
matching which does not address the question of whether disguise impairs performance
compared to performance for recognising across different undisguised images of a face.

1.6 Overview of Current Work
There has been a great deal of previous work on human face-matching performance,
however these studies have generally looked at performance in tasks where the models
used to create same identity stimuli have cooperated with the identity effort and
different person image pairs have been constructed from a small number of available

44

images. Performance could be even worse for more challenging stimuli. There has been
some past work on the challenging case of deliberately disguised faces, but this has been
limited to investigating on the case of evasion disguise, generally through purely the
addition of props to a face.

This thesis will investigate face-matching performance for challenging images, including a
more thorough examination of deliberate disguise. Past research has established that
people perform poorly when matching unfamiliar faces, and familiar viewers perform
very well even when images are of degraded quality. These findings come from
experiments that have tested performance accuracy for matching cooperative stimuli. In
reality, there are situations when people naturally look a lot like somebody else, and also
situations where people look naturally different across multiple images of themselves.
The images that result from these instances of natural similarities between identities or
natural differences in appearance of own identity, would create a far more challenging
task than the cooperative stimuli` tasks created to date. Further to these naturally
occurring instances, people can make deliberate attempts to evade their own identity or
to impersonate another identity. The effect of both incidental and deliberate changes in
own appearance, and incidental and deliberate similarities with someone else’s
appearance will be tested in this thesis. Ways to improve performance will also be
investigated with the effect of familiarity on face matching performance remaining a key
theme throughout. Image manipulations and performance on face-matching tasks for
challenging stimuli will help to address theoretical questions regarding the methods used
for identity judgments.

The second chapter of this thesis explores whether face-matching accuracy is even poorer
for challenging stimuli (ambient same identity images, and very similar different person
image trials) than the poor performance found previously for matching tasks based on
cooperative stimuli. To increase task difficulty and match the challenging image
conditions often encountered in forensic investigations, images are degraded in quality

45

through image pixelation across a series of three experiments. The effect of familiarity on
task performance is explored and treated as a graded rather than binary variable.

Chapter three acknowledges the problem of reducing image quality on face-matching
accuracy that is established in Chapter two, and investigates ways of improving
performance. Three techniques are tested, i) image manipulations; ii) data analysis; and
iii) observer effects. These techniques are first investigated in terms of any improvement
gain they bring when used alone. The effect of combining methods is also examined.

Chapter four investigated whether camera-to-subject distance changes influence face
recognition, with an aim of investigating theory behind facial recognition, whether facematching performance is impaired, and finally whether perceptual constancy methods
exist to facilitate identification across distances. The chapter tests the effect of changing
camera-to-subject distance on the facial configuration of a face as measured from an
image, and also whether any changes result in difficulties in identity judgments. This is an
incidental appearance manipulation, and may make the same identity look different
across multiple images, and perhaps even different identities appear even more different.
Performance accuracy for both familiar and unfamiliar viewers is again tested and
perceptual constancy to help deal with distance induced changes to a face is explored.

The final chapters directly address deliberate changes in appearance through the creation
of a new disguise face database, which encapsulates both evasion and impersonation
disguises and also non-disguised images of these faces. Chapter five tested the effect of
disguise on face matching performance and established effects of disguise type. The
effect of familiarity was investigated and also whether unfamiliar viewers performance
accuracy could be improved if they were informed that disguises might be present.
Chapter six is an exploratory chapter, which aims at further understanding what people
do to create evasion and impersonation disguises, and which of these approaches are

46

effective. This chapter concludes with an experiment that tests for differences in
perceived personality judgments across non-disguised and disguised faces.

Taken together, performance accuracy for the challenging stimuli investigated is even
worse than performance in previous face matching tasks that were constructed from
cooperative stimuli. There are important distinctions between natural and deliberate
efforts to not look like oneself and to look like another person. This is particularly evident
in the study of deliberate disguise. Evasion and impersonation disguise cause different
levels of face matching difficulty and the disguises themselves are achieved through the
use of different methods. Familiarity aids performance in all tested challenging facematching tasks, and there are several other methods that can be used to improve
performance which do not rely on familiarity with the faces concerned.

47

Chapter 2 – Familiarity & Challenging Faces
2.1 Chapter Summary
This chapter looks at the role of familiarity in a naturally difficult identification task –
specifically matching similar faces. In Experiment 1 a graded familiarity advantage is
reported, with participants being poor at the matching task for unfamiliar faces and much
better for familiar faces. This graded familiarity advantage survived even when the images
were pixelated to eliminate fine scale information in the images (Experiment 2) but began
to break down under coarse pixelation (Experiment 3).

Pixelation makes featural and configural information difficult to access. The observed
advantage of extremely familiar faces even under coarse pixelation suggests that other
information besides fine scale information in the images was being used to support the
required discriminations.

2.2 Introduction
Facial identification is a task people often assume they are good at (Bruce, 1988; Jenkins
& Burton, 2011). This belief is likely held because our everyday experiences of face
recognition with familiar faces cause us little difficulty. People can very easily identify
family members, friends, and celebrities across a wide range of image conditions
including different angles, different lighting, changes in pose and even over images taken
years apart (Bruce, 1982; 1994; Jenkins, White, Van Montford & Burton, 2011).
Confidence in human face recognition ability is so high that facial identification forms the
basis of many major security systems worldwide. For example, passport control verifies
personal identity by comparing photographic identification documents against the face of
the holder (physically present).

48

Experimental evidence suggests that this confidence is misplaced. Decades of research
has shown that people make frequent errors in face identification tasks when the tasks
involve unfamiliar faces. This is problematic as the security situations relying on face
recognition as an identity verification method generally involve unfamiliar faces, not
familiar faces. Experiments have tested human face-matching performance in a variety of
ways, including paired face-matching tasks, line up arrays and live person to photo
matching (Kemp et al. 1997; Bruce et al. 1999; Burton, White & McNeil, 2010).
Performance on all of these tasks has been found to be highly error prone, with people
making around 10-20% errors in identity judgments when matching faces, depending on
the details of the task (Burton et al. 2010, Bruce et al. 1999 & Kemp et al. 1997).
Furthermore, experience and training in face recognition appears to make no difference
to task performance - passport officers performed no better than a random sample of
undergraduate students in a recent face identity task that mimicked the passport control
face-matching procedure (White et al. 2014).

The fact that people are making around 1 in 5 errors in such tasks is particularly worrying,
given that this performance level is from tasks where people (or photograph face stimuli)
viewed in the tasks, have been cooperating with the procedure, that is, they have not
been trying to subvert the identification (Henderson, Bruce & Burton, 2001; Burton et al.
2010). The stimuli for face recognition experiments are often new, controlled
photographs taken specifically for the study. For instance, face photographs are usually
taken under consistent lighting, using good quality cameras, posing a neutral expression
and captured from a front on angle. When several different images of the same person
have been taken for a study, there has been no deliberate attempt to make images of the
same person look different across these multiple photographs. Same person photographs
have mostly been captured within very short time intervals – often only minutes apart –
reducing natural variance in appearance due to change in style or age. It is thus very
possible that performance on such tasks reflects a level of performance that is higher
than would be achieved in less favourable but more realistic conditions that include
incidental image variation. For example, current legislation allows a passport photograph
to have been taken up to ten years prior to use. Matching across this decade span is likely

49

far more challenging than an experimental task for which all photographs were taken just
several weeks before testing (White et al. 2014). Similarly, the image quality available
from CCTV could be very poor compared with images in laboratory studies.

Another characteristic of previous face-matching studies is that they have drawn their
mismatch identities from a rather limited pool. Different person trials have typically been
created by pairing together people whose face images look a bit similar to each other –
perhaps due to similar hair colour or face shape. All false match images have to be
selected from the available pool of photographs in the stimulus set. Images used in
mismatch trials are thereby not always convincingly similar in appearance, even for
unfamiliar viewers. This could make them easy to reject.

The upshot is that face-matching ability for very similar faces could be even worse than in
previous studies. In certain applied situations, for example, when using a fraudulent
passport, people may have very strong incentives to use the identity of a person who
looks naturally similar to them in appearance. Yet little is known about viewers’ ability to
discriminate highly similar faces.

Familiarity
Familiarity has been shown in past studies to predict performance accuracy for both same
person and different person identification (Burton et al. 1999, Jenkins et al. 2011). Given
the influence of familiarity on past studies of face recognition, familiarity may be an
important factor in face-matching where tasks include naturalistic same person image
pairs which encompass natural within person variation, and also different image pairs
which are of extremely similar appearance.

50

Experimental participants regularly perform at ceiling level in face-matching tasks when
the images available for comparison are of faces that are familiar to them (Hancock,
Bruce & Burton, 2000). For example, highly accurate performance is achieved even when
the image shown at testing differs from that shown at initial presentation in facial
expression or in the photographed angle (Bruce, 1982). Jenkins et al. (2011)
demonstrated that the familiarity effect holds strong when many photographs are
compared and when the images available for comparison are uncontrolled, and highly
varied on conditions including pose, lighting, expression, hairstyle, and age. In one of their
studies, participants were presented with 40 shuffled face picture cards. Unknown to the
participants, this card deck comprised of 20 face picture cards of one female Dutch
celebrity, and the other 20 face picture cards were of another female Dutch celebrity.
Both unfamiliar viewers (20 British participants) and familiar viewers (20 Dutch
participants) were asked to sort the card deck by identity, grouping together photographs
which they believed showed the same individual’s face. Participants were given no time
restriction and could make as many or as view groupings as they felt reflected the
number of identities present. A strong familiarity effect was found in this experiment.
Unfamiliar viewers struggled with the task, dividing the deck into 7.5 identity piles on
average. But familiar viewers easily group the cards into the two correct identities.
Familiarity with the faces concerned made the task easy, even though the stimuli were
highly variable and this familiarity advantage extends to poor quality images (Bruce et al.
2001, Burton et al. 1999).

Findings such as these highlight that familiarity can make face identity decisions easy,
even when the same decisions are difficult for unfamiliar viewers. However, there has
been very little research into whether familiarity helps with distinguishing extremely
similar faces, such as the face of a disguised imposter from the true identity. Some
support for the notion that familiarity could help comes from studies on telling twins
apart. Stevenage (1998) found that after corrective feedback training, participants rated
photographs of the same twin to be more similar, and images of different twins as less
similar. Robbins & McKone (2003) also report training participants to distinguish identical
twins, also using corrective feedback to aid learning. This study focussed primarily on

51

holistic processing rather than identity judgment. Although these studies did not test the
familiarity effect directly and in isolation (as they gave corrective feedback between trials)
the finding that identical twin faces can be learnt and distinguished after much exposure
to them during the training phase provides scope for a potential advantage for familiarity.
As familiarity helps in the case of distinguishing the faces of identical twins, there is
reason to think that familiarity with a face provides a good starting point for investigating
performance in a matching task involving unrelated, but very similar faces.

Familiarity: a Graded Effect
There are problems in defining familiarity with a face. Experimentally, face groups have
often been divided into familiar and unfamiliar faces for all participants – for example
celebrity faces as the familiar face set (Clutterbuck & Johnston, 2002) and a convenience
sample of non-celebrity faces as the unfamiliar set (Burton et al. 2010) or faces of
celebrities from other countries who are unfamiliar to the participants being tested
(Burton, Kramer & Ritchie, 2015). An alternative approach has been to create a familiar
face set based on colleagues or classmates of the experimental participants and to
compare their performance with another group of participants who would be unlikely to
know the target faces (Burton et al. 1999). Still other studies have familiarised viewers
with novel faces as part of the experiment (Clutterbuck & Johnston, 2004, 2005). These
methods are not without their faults. For example, not all faces that are presented as
familiar (usually celebrity faces) are familiar to all participants. Even among the faces that
are known to the participants, it is unlikely they will be equally familiar to them. For these
reasons, a face cannot always be neatly categorized as familiar or unfamiliar to a single
participant, let alone to a group of participants.

Clutterbuck & Johnston were among the very first to demonstrate the graded nature of
familiarity experimentally (Clutterbuck & Johston, 2002; 2004; 2005). In the first paper in
this series, the familiar and moderately familiar faces used in the study were all celebrity
face images, with each celebrity’s familiarity category chosen on the basis of familiarity

52

ratings provided by eight independent raters. An increase in familiarity led to a decrease
in the time taken to match a full-face image to images showing just the internal features
of a face (see Figure 2.1). Participants were fastest at matching a picture of the full face to
images of internal features for highly familiar faces, slower for moderately familiar faces,
and slowest for unfamiliar faces (Clutterbuck & Johnston, 2002). There was however no
significant difference in performance when matching full-face images of each of the
categories (highly familiar, moderately familiar, unfamiliar) to same or different images of
the external features of the face.

Figure 2.1 Example of full face and internal feature stimuli (left) and full face and external features (right)
viewed as part of Clutterbuck & Johnston’s (2002) face-matching task.

Clutterbuck and Johnston (2002) assumed that the faces in each of the categories were of
the same familiarity level for all participants (i.e. a face in the highly familiar category was
assumed to be highly familiar for all participants). However, the raters who rated the
familiarity of the face, did not take part in the experiment itself, so the familiarity bands
may not be an accurate reflection of how familiar each face was to each of the
participants in the main study. It is possible that some participants would be more
familiar with the moderately familiar celebrities and vice versa.

53

Clutterbuck and Johnston (2004; 2005) carried out two additional studies that address
these concerns to some extent. They used celebrity face images for the familiar category,
newly learnt faces for mid-level familiarity and novel (previously unseen) faces to provide
the unfamiliar level. The graded familiarity advantage was also observed using this
method of familiarity division. Performance improved as familiarity increased on tasks
involving gender judgement (Clutterbuck & Johnston, 2004) and face-matching speed
(Clutterbuck & Johnston, 2005). Taken together, these findings strongly support
familiarity as a continuous variable rather than a binary concept – people can be more or
less familiar with a face, and this level of familiarity will affect performance on facematching tasks in a graded way.

When investigating the effect of familiarity on face-matching ability for very similar faces,
and using images of same faces that incorporate natural variation, it will be important to
ensure that familiarity is measured in a way that accounts for i) familiarity being a graded
concept and ii) the notion that not all celebrity faces will be equally familiar to all viewers.

2.3 Celebrity Faces & Celebrity Lookalikes
Testing for a familiarity advantage requires faces that differ in their degree of familiarity.
Most people are highly familiar with their friends, colleagues and family members. The
problem with this is that familiarity with these faces is very specific to the few individuals
who know them. A study on face familiarity requires a large number of images of faces
that will be familiar to many people who view them in the study, and also a large number
of images of faces that are unfamiliar to these people. Ideally each of the faces used
across the experiment will be familiar to some participants but unfamiliar to others, and
all of the faces will be familiar to at least some of the participants. Celebrities provide a
group of identities who are familiar to a very large array of people, and celebrity face
images are easily accessible via Internet search. An additional advantage of using celebrity
images is that there are many different categories of celebrities including pop stars,
reality television stars, actors, politicians and sports personalities. This range gives the

54

freedom to choose images of celebrities from very different settings. Based on personal
interests, each participant will be more or less familiar with celebrities from each of these
categories. If celebrity images are sampled from a wide range of categories (singers,
actors, politicians etc.), it is likely that each participant will be familiar with at least some
of these celebrity faces and less familiar with others.

Use of Lookalikes as a Proxy for Imposters
In order to investigate performance for a challenging face-matching task proposed, it is
necessary to have access not only to different photos of the same face that vary naturally
in their appearance - such images are known as ambient images (Jenkins et al. 2011;
Sutherland et al. 2013) - but also to photos of other faces that look extremely similar to
the targets.

Conveniently, there is a ready source of faces that closely resemble celebrities and that is
the celebrity lookalike industry. In the following experiments, I will use celebrities and
their professional lookalikes to construct highly similar face pairs.

2.4 Experiment 1: Lookalike Task
The purpose of Experiment 1 was to test the effect of familiarity on performance accuracy
in a challenging face-matching task. All of the face stimuli are ambient images to sample
the natural variation in each person’s appearance. This will presumably result in more
challenging same person trials than in past work that has used highly controlled
cooperative images. The use of celebrity lookalikes as imposters in my experiment should
allow for extremely difficult different-identity pairs, compared with those used in
previous experiments. The intention here is to model real life situations where someone
may be trying to pass impersonate a similar looking person on a fraudulent security
document.

55

If familiarity helps with these very fine distinctions, then matching performance should be
more accurate for increasing levels of familiarity.

Method
Participants
30 undergraduate students (M = 8, mean age = 20.2) at the University of York
volunteered as participants for this project. All participants were paid £3 or a half hour
course credit in return for their participation.

Stimuli
Image Selection
Face images of 30 celebrity identities (three face images per celebrity), and one face
image of a professional lookalike for each of these celebrities, were selected as
experimental stimuli (120 images in total). This number of images was necessary to
provide two celebrity face photographs to constitute the same-person pair, and one
additional celebrity face image to be paired with the lookalike image to create the
different pair in the face-matching task. For a celebrity to be included, they needed to
have at least one professional lookalike whose image was accessible from the Internet.
For a lookalike image to be chosen as suitable for use in the study, a viewer who was
extremely familiar with the celebrity in question approved their high level of visual
similarity to the celebrity.

Following approval of the celebrity and lookalike images in the image selection, two face
image pairings were created for each celebrity – one showing two different images of the
celebrity, and the other showing a third image of the celebrity paired with an image of
that celebrity’s lookalike (see Figure 2.2). The lookalike image could appear on either side

56

of the screen, and appeared on the left and right side equally often across the
experiment.

Figure 2.2 Face-matching task image examples. The pairs on the left (A) show different identities (with the
imposter face on the right), the pairs on the left (B) show same identity pairs.

Design
This experiment was conducted using a within subjects design, which tested the effect of
five levels of face familiarity on the dependent variable, which was performance accuracy
in the face-matching task.

A novel familiarity scale was designed for use at the end of the experiment to assess how
familiar each of the celebrities was to each of the participants. This scale was set across
the desk where the experiment took place and was a meter long in length, marked for 0100cm, with 0 representing unfamiliar faces and 100 representing faces that were
extremely familiar (see Figure 2.3). This scale was intended to address two limitations of
previous familiarity manipulations, i) not all familiar (celebrity) faces are familiar to
everyone, ii) some faces will be better known than others and this is the case for each
participant.

57

Procedure
Face-Matching Task
Participants took part in a face-matching task involving 60 image pairs (two pairs for each
of the 30 celebrity identities), viewed on a computer screen. The participants’ task for
each pair was to determine whether the two images showed the same identity or
different identities (i.e. one of the images was of the lookalike). Two different random
orders of image pair presentations were created; each participant was assigned to view
one of these two random orders. Participants were informed that the lookalike images
could appear on either side of the screen and that there was no time limit for completing
the task.

Familiarity-Rating
On completion of the face-matching task participants were given photograph cards (size
6cmx4cm) of each of the celebrities that they had seen in the face-matching task (N = 30)
and asked to rate them for how familiar they were with the celebrity’s face before
completing the task. Participants received just one of the three true celebrity images that
they had viewed in the face-matching task for use in the familiarity-rating task. The image
viewed for each celebrity was selected randomly from the three available, and the chosen
card for each celebrity remained the same for all participants. Participants rated the faces
for familiarity by placing them on a scale that ran from 0 (completely unfamiliar) to 100
(extremely familiar). Faces of equal familiarity could be placed down vertically one above
the other to create a column of equally familiar faces; this was particularly useful for faces
that were completely unfamiliar and extremely familiar (see Figure 2.3 for image of the
familiarity scale in the experimental setting). This approach allowed me to capture the
relative and absolute familiarity of the celebrities separately for each participant.

58

Figure 2.3 Photograph of one participant’s use of the familiarity scale taken immediately after completion in
the experimental setting. The far left side of the scale indicates that the face was completely unfamiliar, and
the far right depicts extreme familiarity with the face.

Analysis
The main measure of interest was the percentage of correct responses in the facematching task. To examine the effect that familiarity score had on percentage accuracy in
the face-matching task, I grouped the raw familiarity ratings into 5 different familiarity
levels. This was achieved by binning each participant’s face matching responses into 5
familiarity bands (quintiles) based on the participant’s own ratings. In total 35% of faces
were placed in Band 1, 9% in Band 2, 10% in Band 3, 8 % in Band 4 and 39% in Band 5.
Interestingly, although the majority of faces were placed at the extremes of the spectrum
(0-19 and 80-100) all participants placed some faces in the middle familiarity bands (2039,40-59, 60-79).

Results
Percentage accuracy scores were entered into a one-way repeated measures ANOVA to
investigate the effect of familiarity on face-matching performance in this celebrity versus
lookalike discrimination task. Percentage scores were rounded to integers throughout.

59

Figure 2.4 Percentage of correct responses in face-matching task (using fine quality 200x300 pixel images)
for each familiarity quintile; 1 (0-19), 2 (20-39), 3 (40-59), 4 (60-79), 5 (80-100). With Band 1 being
completely unfamiliar and Band 5 being extremely familiar. Error bars show standard error of the mean.

The results revealed that face-matching performance was significantly affected by the
Familiarity with the face viewed, F(2.98, 86.51) = 10.07, p = <.001, η p2 = .26 (GreenhouseGeisser corrected). Accuracy was lowest for the faces that were most Unfamiliar (Band 1,
M = 72%, SE = 1.58, CI = 69 -76), and best for the faces that were most Familiar (Band 5,
M = 90%, SE = 1.95, CI = 86-94). As can be seen in Figure 2.4, there was a generally graded
increase in performance as familiarity increased (Band 2: M = 70%, SE = 2.98, CI = 64 -76;
Band 3: M = 77%, SE = 2.94, CI = 75-88; Band 4: M = 81%, SE =1.95, CI = 77-84).

60

Figure 2.5 Pairwise comparisons showing which familiarity levels performance was significantly better than
the other familiarity levels.

Pairwise comparisons revealed that the performance was significantly better for faces
that were placed in familiarity Band 5 than for faces in Bands 1,2 and 3 but not 4 (see
Figure 2.5).

61

Figure 2.6 Graph showing pattern for Same identity pairs correct response and Different (lookalike) identity
pairs correct responses. Error bars show standard error of the mean.

I then analysed accuracy separately for Same identity and Different identity trials (Figure
2.6). For both types of trial accuracy increased as familiarity increased. As expected there
was somewhat higher accuracy in the Same identity condition than in the Different
identity condition, presumably reflecting a tendency to judge highly similar faces as the
same person. A 5x2 within-subjects ANOVA with the factors of Familiarity and Pair Type
confirmed that this difference was significant: overall accuracy was significantly higher for
Same trials (M = 81%, SE = 2.11, CI = 77-85) than for Different trials (M = 74%, SE = 1.84, CI
= 71-78), [F(1,29)=5.31, p<.05, η p2 =.23). As expected, there was a main effect of
familiarity [F(4,116)=11.10, p<.001, η p2 =.12], indicating a graded familiarity advantage for
Same and Different identity trials alike. There was no interaction between Familiarity and
Pair Type – familiarity had no more of an effect for correct performance on Same identity
trials than Different identity trials [F(4,116)=.38, p=.83, η p2 =.01].

62

Discussion
Participants were able to make remarkably fine discriminations between extremely
similar faces (celebrities and their lookalikes), and to integrate naturally varied same
identity images. This ability was underpinned by a graded effect of familiarity. Accuracy
was at its peak when the faces viewed were extremely familiar to the participants
(performance accuracy 90%), yet performance was much worse for the completely
unfamiliar faces (accuracy 72%, note that chance performance is 50%). Task performance
generally increased with a progression in familiarity (Band 2 M = 70%, Band 3 M = 77%,
Band 4 M = 81%). Importantly these results are based on personal familiarity scores for
each of the celebrities. This method of analysis allowed a graded familiarity effect to be
teased out despite the fact that not all participants were familiar with the same
celebrities.

Performance for completely unfamiliar faces in the lookalike study (72% accuracy) was
lower than identity judgment accuracy found in past studies (e.g. Bruce et al, 1999;
Megreya & Burton, 2007; Burton et al. 2010). When compared to performance on the
GFMT, a standardised test of face-matching ability that comprised of cooperative stimuli
and limited false match image options, mean performance accuracy is lower on the
lookalikes task. This is the case for overall performance accuracy on the unfamiliar face
images (lookalikes task mean performance =72%, GFMT = 89.9% (long version), 81.3%
(short version), and also for same identity trials (lookalikes = 74%, GFMT = 92% [long
version], 79.8% [short version]) and different identity trials (lookalikes = 70.7%, GFMT =
88% [long version], 82% [short version]). Strikingly, performance is nearly 20% worse for
the naturally varied images of the same celebrity faces in my study, than the cooperative
same person images used in the GFMT. Furthermore, performance is more than 10%
worse for different person trials in my lookalikes task than even in the short version of the
GFMT, which includes only the hardest items in the GFMT. This highlights that both
photographs of the same person which include natural variation, and very similar looking
different identities, cause more problems to face matching performance than previously

63

captured by tasks constructed from cooperative stimuli. This is concerning given that both
of these image types may be encountered in security situations and attempts of fraud.

As in past face recognition research which has shown a familiarity advantage for both face
memory and face-matching (Bruce, 1986, Ellis et al. 1979, Burton et al. 1999), familiarity
had a great impact on face-matching performance. The present study shows that the
benefit of familiarity extends to the challenging case of extremely similar lookalike faces,
and naturally varying images. Whereas past studies were often very easy for familiar
viewers, resulting in ceiling performance (Hancock, Bruce & Burton, 2000), the lookalike
task brought performance off ceiling, so that modulations in performance could be
observed. Even with these very challenging viewing conditions accuracy reached 90% in
the highest familiarity band. Thus, it seems that the effect of familiarity is so strong, that
even professional lookalikes are an unconvincing false match for viewers who are
extremely familiar with that celebrity.

My study broke down familiarity even further than in previous studies, by comparing 5
levels of familiarity rather than just 3 as used by Clutterbuck & Johnston (2002, 2004,
2005). Assessing performance across 5 levels of familiarity provided a more realistic and
accurate categorisation of participants’ familiarity with the faces, allowing a more
detailed exploration of the extent of graded effect of familiarity. Familiarity with a face is
not an all-or-nothing phenomenon. Different viewers are familiar with different faces to
differing degrees. Constructing the analysis around that insight reveals much finer
structure than a binary familiar/unfamiliar distinction allows.

Now I have demonstrated that the graded familiarity advantage survives to the
challenging case of imposter detection, I want to see how much further I can push the
effect of familiarity. So far, I have only tested the effect for fine quality (un-manipulated)
images. Although I have found that familiar viewers can tell apart a target face and
imposter, it is not clear what they have learnt about the familiar face which has allowed

64

them to do this. Given that the faces in the task were very similar (i.e. celebrities and their
lookalikes), it seems likely that fine scale information in the image is critical for making
the necessary discriminations. If so, then obscuring the fine-scale information should
impair performance, even for familiar viewers. There are several possible techniques for
obscuring image detail. Here I used image pixelation, whereby the number of pixels in the
image is reduced while image size is held constant. This technique has an interesting
pedigree in the psychology literature (Harmon & Julesz, 1973). It also arises in the context
of applied face identification whenever digital images are enlarged (Jenkins & Kerr, 2013).

2.5 Experiment 2: Mid Pixelation
In Experiment 1 a graded effect of familiarity on face matching performance was evident
for both integrating different images of the same face and telling apart very similar faces
of different identities. In that experiment, fine quality images were used as stimuli.
However, face images encountered in applied matching tasks are often not of good
quality. For example, coarsely pixelated images may be obtained by zooming in on Closed
Circuit Television (CCTV) footage to gain an image of a suspect’s face (see Figure 2.7 for
applied example).

Figure 2.7 Example of actual image issued by the police to the public to assist with identification of a man
caught on CCTV (Howarth, 2016). This image takes a pixelated appearance.

Face recognition for unfamiliar faces is poor even under favourable image conditions (see
Burton et al. 2010). The task is even more challenging when the images are of low

65

resolution as pixelation disrupts the information that is available from an image (Harmon
& Julesz, 1973). When an image is pixelated, the information from several adjacent pixels
is pooled to form larger pixels. The luminance of the new pixel is determined by averaging
together the luminance values of the constitute pixels. High spatial frequency information
within that area is thus lost. In tandem, high spatial frequency noise is introduced at the
edges of the new pixels. This is due to the larger changes in luminance between adjacent
pixels in the new image than between adjacent pixels in the original image. This noise is
particularly disruptive to viewing, as the visual system is highly sensitive to lines and
geometric patterns, making the new pixel boundaries difficult to ignore (Harmon & Julesz,
1973).

As a result of this manipulation, it is difficult to extract exact information about a face
from a pixelated image. In particular, when we view a pixelated face image, configural
information (the metric distances between features) that can be extracted from a face
becomes less precise, and the appearance of features becomes less detailed.
Unsurprisingly, the pixelation manipulation has been shown to increase the difficulty of
image recognition, compared to a non-manipulated version of the image (e.g. Harmon &
Julesz, 1973, and replicated by Morrone, Burr & Ross, 1983; Sergent, 1986; Bachmann,
1991; Costen, Parker, & Craw, 1994, 1996; Uttal Baruch & Allen, 1995). However, people
can name pixelated celebrity images, even at very low levels of pixelation (Lander, Bruce,
& Hill, 2001). The work of Lander and colleagues (2001) focused on pixelation from the
perspective of identity protection issues – images are often pixelated by the media in
order to try and protect identity. They argued that pixelation is not an effective method of
obscuring identity, as the face remains identifiable to a familiar viewer. Although
recognition ability (being able to say who the face is) for pixelated images of famous or
learnt faces has been tested in the past, face-matching (determining whether pairs of
images depict the same person or different people) with pixelated faces has received little
attention.

66

Bindemann, Attard, Leach & Johnston (2013) were the first to test performance in a
matching task using pixelated unfamiliar faces. In their study participants were tasked
with matching a pixelated image to a good quality image. The stimuli used in this study
were the images used in the GFMT, where one image in each of the image pairs had been
pixelated (see Figure 2.8). They found that participants were much better at matching
two intact face images, than matching one intact face image to a pixelated image.
Moreover, performance declined as the level of pixelation increased. However, in some
practical situations investigators may have to compare multiple pixelated images from
different CCTV footage to try and piece together an event sequence. I have been unable
to find any published research that previously tested such performance. In addition to
this, Bindemann et al. (2013) used staged face photographs, using either posed profile or
front view face image (Burton et al. 2010). For example, images captured from CCTV
footage may vary greatly in lighting and pose. It is unlikely that such photographs would
be available in forensic settings. For these reasons, it is important to measure facematching ability for pixelated images using ambient images, as past research has shown
that the same face can appear drastically different between ambient photograph images
(Jenkins et al. 2011).

Figure 2.8 Example of stimuli used in the face-matching task created by Bindemann et al. (2013)

These issues will be addressed by Experiment 2, which will test people’s ability to match
pairs of pixelated ambient images. This new task is more difficult than previous
recognition and naming tasks because both the lookalike and the celebrity image map

67

onto the same individual. For example, the viewer might associate both images with say,
Al Gore, but that is not enough to solve this task. The problem is to decide whether both
images actually show Al Gore, or whether one of the images shows an imposter
(lookalike). Additionally, the nature of the paired matching task gives a baseline score.
Unlike in a naming task, chance performance is known to be 50% in the paired matching
task, so observed performance can be compared against this chance level. Finally, the
ambient images used in this task provide us with a test more similar to the image type
available in real world investigations.

To address all of these issues I repeated Experiment 1, but this time replacing the fine
quality ambient images with pixelated versions of these images. The aim is to establish
whether the graded familiarity advantage observed for imposter detection in fine quality
images extend to degraded images. If it is knowledge of exact facial configurations and
fine featural detail that differentiates familiar and unfamiliar viewers in the lookalike task,
then obfuscating that information should eliminate the familiarity advantage.

If familiar viewers are using the detail and small differences in faces to solve the task,
then there should be little or no benefit of familiarity in face-matching performance in
Experiment 2, where this information is difficult to access. On the other hand, if the
familiarity advantage survives, that would suggest that other information is being used.

Method
Participants
30 undergraduate students at the University of York (M = 11, mean age = 19.7)
volunteered as participants for this project. Participants received payment of £3 or a half
hour course credit. None of the participants had taken part in the previous experiment.

68

Design & Stimuli
As in Experiment 1, a within-subjects design was adopted to compare the effect of
Familiarity (5 levels) on the dependent variable, percentage of correct responses in the
face-matching task.

The image pairings of the 30 celebrity faces and their 30 lookalikes faces were the same
as Experiment 1. However, unlike Experiment 1, all of these images were pixelated to a
level of 30 pixels wide 45 pixels high using Adobe Photoshop (CS6) (see Figure 2.9 for a
side by side example of the stimuli used in Experiment 1 and the pixelated versions for
use in Experiment 2).

Figure 2.9 Example of the image appearance for Experiment 2 (top pair) compared with the fine version of
the same image as used in Experiment 1 (bottom) pair. These are different image pairs of Al Gore with the
lookalike appearing on the right.

69

Procedure
The procedure was the same as in Experiment 1, except that this time participants viewed
the newly created pixelated versions of the image pairs.

Following completion of the face-matching task, participants ranked the celebrity faces
for familiarity using the familiarity scale as in Experiment 1. Full resolution (unmanipulated) images were used for the familiarity-ranking task as in Experiment 1.

Analysis
As in the previous experiment, results were analysed by comparing the percentage of
correct responses across the familiarity quintiles. In this experiment 34% of faces were
placed in Band 1, 12% in Band 2, 9% in Band 3, 10% in Band 4 and 34% in Band 5.

Results
A one-way repeated measures ANOVA was performed on the accuracy data, to
investigate the effect of familiarity (5 levels) on participants’ face-matching ability for
poor quality images.

70

Figure 2.10 Graph showing the graded effect of familiarity for participants’ face-matching task
performance. Error bars show standard error of the mean.

As in Experiment 1, face-matching performance was significantly affected by Familiarity
with the face viewed, F(3.19, 92.48) = 2.96, p<.05, η p2 = .09 (Green-House Geisser
corrected). Figure 2.10 shows the predominantly graded effect of Familiarity over the 5
familiarity bands (Band 1: M = 64, SE = 2.08, CI = 59 -68; Band 2: M = 66, SE = 3.22, CI =
59-73; Band 3: M = 70, SE = 2.96, CI = 64-76; Band 4: M = 68.69, SE = 4.07, CI = 60-77;
Band 5: M = 76, SE=2.79, CI = 70-82).

Pairwise comparisons revealed that accuracy for the highly familiar faces (Band 5) was
significantly better than performance for faces in familiarity Band 1, mean difference =
12.28, SE = 2.75, CI = 6.65-17.91, p<.005 and also than that of Band 2, mean difference =
9.75, SE = 3.96, CI = 1.64-17.85, p<.05. There were no significant differences in

71

performance accuracy between each of the other familiarity bands (p>.05 for all
comparisons).

Figure 2.11 Face-matching performance broken down into correct Same and Different identity trials. Error
bars show standard error of the mean.

Accuracy data were also analysed according to the breakdown of Same identity and
Different identity correct trials using a 2x5 ANOVA (see Figure 2.11). A significant main
effect of Familiarity was observed [F(4,116) = 3.00, p<.05, η p2 =.09]. There was also a
significant main effect of Trial Type - participants were overall more accurate at the Same
identity trials than the Different identity trials [F(1,29) = 4.99, p<.05, η p2 =.15]. However,
there was no significant interaction between Familiarity and Trial Type [F(4,116)=.79,
p=.53, η p2 =.03] (see Figure 2.11).

72

Discussion
Despite the pixelation of the images, accuracy in the lookalike task improved as familiarity
increased.

My pixelated lookalike task was naturally more challenging than previous identification
tasks involving pixelated images. Firstly, the lookalike matching task is more challenging
than naming pixelated celebrity images (Lander et al., 2001), as in my task both the
lookalike and celebrity lookalikes could be mistaken for (and named as) being the same
celebrity. In addition to this, previous matching tasks have involved matching one good
quality image to one pixelated image (Bindemann et al. 2013), whereas my task required
matching across two pixelated images. The findings from my challenging pixelated
lookalike experiment hence demonstrated the versatility and strength of familiarity as an
aid to face recognition, as even though the celebrity and the lookalike images were
pixelated, the graded familiarity effect on face-matching performance prevailed. Finally,
this study suggests that for familiar faces, learnt information other than featural details
and configural information may be used to perform the task. Fine-scale image
information was more difficult to access than in Experiment 1, yet the familiarity
advantage survived despite this.

At the current level of pixelation (30x45 pixels) some featural and configural information
is still visible. This raises the question of where the familiarity advantage will break down.
Presumably in the limiting case (1x1 pixel), performance on this task would be at chance
for all familiarity bands. The graded familiarity effect in the present experiment implies
that viewers were nowhere near their performance limit in this task. In the next
experiment, I set out to push the familiarity advantage to its limits by pixelating the
images even further.

73

2.6 Experiment 3: Coarse Pixelation
Past research has shown that there is a limit to people’s ability to recognise a pixelated
face – we can recognize pixelated faces but only up to a point. The lower the image
resolution the higher the error rate, with results eventually falling to chance level
(Bachmann, 1991; Costen et al., 1994, 1996). It is believed that a familiar face can be
recognised up to horizontal pixelation level of 16 pixels per face, any level beyond this
results in a steep decline in performance (Bachmann, 1991; Costen et al., 1994, 1996).
Bindemann et al. (2013) reported that the pixelation threshold was in fact much lower for
matching unfamiliar faces. Bindemann et al.’s (2013) study reported limits on
participants’ abilities to match two side-by-side images, i.e. the scenario in the lookalike
experiment series. Bindemann and colleagues (2013) reported that a large drop in facematching performance accuracy occurred when one of the high-resolution images in each
of the image pairs was replaced with images of a horizontal resolution of 20 pixels.
However, Lander (2001) found that people could identify around half of the familiar face
photographs presented to them when the images comprised of a horizontal resolution of
only 10 pixels per face. In Bindemann et al.’s (2013) unfamiliar face-matching task
performance was at around chance level for a horizontal pixelation of 8 pixels, even
though the face-matching task consisted of co-operative stimuli (taken from the GFMT)
presented side by side. Thus these previous findings indicate that pixelation will reach a
point where the familiarity advantage no longer holds.

So far the graded familiarity effect has prevailed for ambient images of extremely similar
faces, even when these images were degraded using pixelation. I previously argued that
reducing the number of pixels in the image makes configural and featural information in
the image more difficult to access. The accuracy data from Experiment 2 suggests that
some critical information was still accessible albeit at a reduced level (which could explain
the overall poorer performance in Experiment 2).

74

If the familiarity advantage remains for even coarser pixelation, it would suggest that
information other than fine featural details and exact configurations support high
performance by familiar views in the lookalike task. If the familiarity advantage is
eliminated, this would suggest that the familiarity advantage relied solely on the finescale information that is disrupted by coarse pixelation.

Method
Participants
30 undergraduate students at the University of York (M = 8, mean age = 20.2)
volunteered as participants for this project. Participants received payment of £3 or a half
hour course credit. None of the participants had taken part in the previous experiment.

Design
As in the previous experiments, Experiment 3 was a within-subjects study, which
investigated the effect of familiarity on performance on the lookalike matching task. The
only difference between this experiment and the preceding experiments was the level of
pixelation in the stimulus images.

The pixelation level chosen for this experiment was 20 pixels wide x 30 pixels high. This
particular resolution was selected because Bindemann et al. (2013) reported a marked
drop in performance accuracy for this level of pixelation, although accuracy was still
above chance in their study.

75

Stimuli
The same image pairings of the 30 celebrity faces and lookalike faces for each of these
celebrities were used as in Experiment 1 and Experiment 2. This time the images were
presented at a pixelation level of 20x30 pixels using Adobe Photoshop (CS6).

Figure 2.12 Example of coarsely pixelated image stimuli used in Experiment 3.

Procedure
The procedure for this experiment was the same as for Experiments 1 and 2, except that
the face-matching task now involved the coarsely pixelated images (20x30 pixels).

Following completion of the face-matching task participants used the familiarity scale as
in the previous experiments, to indicate their level of familiarity with each of the celebrity
faces. As in both previous experiments, the good quality image cards were used for the
familiarity judgement task.

Analysis & Results
Experiments 1 and 2 established a graded effect of familiarity on task performance by
examining accuracy at each of five familiarity bands. This approach was not possible for
Experiment 3 because participants used the middle range of the familiarity scale less
frequently. 37% of faces were placed in Band 1, 20% in Band 2 and 43% in Band 3.

76

Dividing the data into familiarity quintiles meant that the middle quintiles (2, 3 and 4)
were too sparsely populated to allow meaningful statistical analysis. To circumvent this
problem and obtain a reliable performance estimate for mid-level familiarity faces, data
from familiarity bands 2, 3 and 4 were pooled into a single band. This resulted in 3
familiarity bands, as used in previous studies (e.g. Clutterbuck & Johnston, 2002, 2004,
2005).

A one way repeated measures ANOVA was performed on the accuracy data, this time
examining the effect of Familiarity (3 levels), on performance accuracy in the facematching task.

Figure 2.13 Percentage of correct responses for each of the three levels of familiarity in the 20x30 pixel
condition. Error bars show standard error of the mean.

Once again, there was an overall effect of Familiarity for this task [F(2,58) = 9.44, p<.001,

η p2 =.25] (see Figure 2.13). Mean performance accuracy was highest for highly familiar

77

faces, Band 1 M =65%, (SE =1.78, CI =62-69). Performance accuracy for Band 2 was M =
54% (SE =2.84, CI = 48-60) and for Band 1 M = 58%, (SE =1.59, CI = 54-60) (see Figure
2.13).

Pairwise comparisons revealed that accuracy was significantly higher for highly familiar
faces (Band 3) than for both faces of mid familiarity (Band 2) (mean difference = 11.22, CI
= 4.67-17.77, p<.005) and unfamiliar faces (Band 1) (mean difference = 7.70, CI = 3.5711.83, p<.005). There was no significant difference between accuracy scores for faces in
familiarity Band 1 and Band 2.

Figure 2.14 Percentage of correct responses in face-matching task (using poor quality 20x30 pixel images)
by familiarity broken down into same (dotted line) and different (dashed line) correct trials. Error bars show
standard error of the mean.

78

A 2x3 ANOVA was conducted to break down results into Same identity and Different
identity correct trials. This analysis revealed a significant main effect of Familiarity
[F(2,58)=6.39, p<.005, η p2 =.18]. There was however no significant main effect of Pair Type
[F(1,29)=1.09, p=.19, η p2 =.04]. There was also a significant interaction between Pair Type
and Familiarity [F(2,58)=6.08, p<.005, η p2 =.17). This is illustrated in Figure 2.14.

There was a significant simple main effect of Familiarity for Same identity trials F(2,116) =
10.29, p<.001, η p2 = .15. There was no simple main effect of Familiarity for Different
identity trials F(2,116) = 2.55, p>.05, η p2 = .04.

As a significant interaction was observed between Familiarity and Pair Type, Tukey post
hoc tests were conducted to find out where significant differences lay. For Same identity
trials, there was a significant difference in performance between familiarity bands Low
(Band 1) and High (Band 3), and between Mid (Band 2) and High (Band 3). There were no
significant differences between any of the levels of Familiarity for Different identity trials.

Discussion
Although a familiarity advantage survived in Experiment 3, with people being significantly
better at recognizing the faces that were extremely familiar to them (M = 65% than the
least familiar face (M = 58%), the graded effect seen in Experiments 1 and 2 did not
emerge here, and accuracy in familiarity Bands 1 and 2 was numerically not much above
chance level (50%). It seems that by reducing image quality to 20x30 pixels we are
approaching the limit of the familiarity advantage in this situation. Breakdown of results
by trial type revealed that this familiarity advantage was driven by improved performance
for same person trials.

79

It is interesting that a familiarity advantage still emerged when comparing extremely
familiar faces to less familiar faces. This advantage suggests a role for coarse scale
information even when discriminating extremely similar faces. Performance was at
around chance level unless the faces were extremely familiar. In some ways it may not be
surprising that there was a performance advantage for highly familiar faces, as it has been
previously shown that people can recognise associates even in very poor quality images
(e.g. Burton et al. 1999). However, the foil faces in previous studies have been generic
similar faces that merely share the same basic description (e.g. young male, short black
hair). The important difference here is that my lookalike foils were themselves
recognisable as the celebrities they were impersonating. The implication is that the
lookalike faces differ from the celebrity faces only in subtle detail. Yet disrupting the
subtle detail in the images was not catastrophic for familiar viewers. This suggests that
the familiar viewers used other information to solve the task. One possibility is that at
least some of the subtle differences are carried in the low spatial frequency information
that is intact in the pixelated images.

Another cue comes from the pattern of breakdown of the familiarity advantage. In
Experiment 3 familiarity did not improve performance on different identity pairs. It is thus
possible that familiarity is making it easier to determine the identity of the target
celebrity from the poor quality image, this identity decision could come from either the
celebrity or lookalike image. When a familiar viewer can identify the celebrity, they then
have access to all the representations that they have stored for that celebrity’s face.
Familiar viewers are aware of many more ways the celebrity’s appearance can take, and
hence allow a greater range of variation of appearances for the face, than an unfamiliar
viewer may. Therefore familiar viewers may be more accepting of saying same to the
image pairs in the matching task, even though the exact details can't be extracted from
the pixelated images. This approach would improve performance for same identity trials
but lead to poorer performance for different person trials, and could thus explain the
pattern of results in Experiment 3.

80

It seems that faces are compared in different ways depending on our familiarity with the
faces involved (e.g. Megreya and Burton, 2006). If people are unfamiliar with the face
they may be matching face images in a pattern matching type manner, similar to the
method used to match images of objects (Hancock et al. 2000; Burton & Jenkins, 2011),
yet when people are extremely familiar with a face, our findings suggest that people no
longer rely on this pattern type matching, but can use other information which we have
learnt for familiar faces, to aid the matching task.

2.7 Between Experiments Analysis
In order to compare the familiarity advantage across experiments (and across image
quality), I next tested how the results of Experiment 3 compared with the results of
Experiments 1 and 2. As familiarity was assessed across three levels in Experiment 3, I
reanalysed the data from data experiments 1 & 2 in the same way to allow direct
comparison.

A 3x3 mixed ANOVA was performed on the results of Experiment 1, 2 and 3. This
compared Familiarity (3 levels) with Image Quality (Experiment, 3 levels [fine, mid &
coarse pixelation]). The ANOVA revealed a main effect for Familiarity, such that facematching accuracy increased with increasing familiarity when pooling over Image Quality,
Band 1, Low M = 64%, Band 2, Mid M = 65%, Band 3, High M = 77%, [F(2, 174) = 35.24, p
<.001, η p2 =.29]. Pairwise comparisons revealed that these differences lay between High
familiarity and both Mid (mean difference = 11.33, CI = 7.78-14.49, p<.001) and Low
familiarity (mean difference = 12.52, CI = 9.81-15.24, p<.001). There was no significant
difference between Low and Mid familiarity, p>.05.

A significant main effect was also observed for Image Quality. Performance accuracy was
highest for the Fine images (Experiment 1) (M=78%, SE=1.65, CI=74.78-81.36) then Mid
pixelation images (Experiment 2) (M=69%, SE=1.65, CI=65.51-72.08), which were both

81

higher than performance for the Coarse pixelation images (Experiment 3) (M=58%,
SE=1.65, CI=54.37-60.94), [F(2,87)= 39.37, p<.001, η p2 =.47].

There was an interaction observed between Familiarity and Image Quality [F(4,174)=
2.94, p<.05, η p2 =.06]. This is due to the graded familiarity effect seen in Experiments 1
and 2 breaking down in Experiment 3 as a result of extreme image degradation.

2.8 General Discussion
It is evident from this series of experiments that being more familiar with a face increases
a person’s ability to tell that face apart from its lookalike, and also to ‘tell together’
ambient images of the same face. Familiarity aids imposter detection even in the case of
poor quality images, with performance increasing in a graded manner when both Fine
pixelation ‘standard’ images were viewed (Experiment 1) and when Mid pixelation images
(30x34 pixels) were used (Experiment 2). However the findings of Experiment 3 illustrate
that the graded familiarity starts to break down as the pixelation becomes more Coarse
(20x30 pixels). Not only does the graded pattern falter in the overall accuracy data, but
also the familiarity advantage is lost for different identity trials.

My findings underscore those of previous research and extend them in several ways. I
created a challenging face matching task that addressed face-matching performance for
ambient celebrity face images and very similar celebrity and lookalike faces. Performance
accuracy on this task was even poorer than had previously been established in matching
tasks which used cooperative stimuli. In my lookalike task unfamiliar viewers performed
with 72% accuracy, which was a level of performance much poorer than in the GFMT (M =
90% long version, 81% short version) which is a standardised face-matching test
consisting of cooperative stimuli (Burton et al. 2010). Accuracy dropped even more with
degraded image quality. These results reflect that human face-matching performance is

82

even worse for challenging images than previously established in standardised
cooperative matching tests.

My research also adds to the existing literature on familiarity as a graded concept
(Clutterbuck & Johnston, 2002, 2004, 2005). Clutterbuck & Johnston divided familiarity
into just three bands (high, medium and low) with faces categorized into these bands
according to ratings from an independent rater group. The current Experiments 1 & 2
provide more detailed insight into familiarity effects by tracking performance over five
levels of familiarity instead of 3. Importantly, in the present experiments, familiarity
ratings were based on participants’ own rankings of their level of familiarity with a face,
unlike in previous studies which have assumed equal familiarity with the familiar face
stimuli for all participants. I also tracked performance across changes in image quality,
and found first that familiarity does help improve performance even for coarsely
pixelated images, but there is a limit to the familiarity advantage, especially its graded
nature. These findings fit with previous demonstrations that face recognition and facematching performance decline as image resolution decreases; with performance
eventually falling to chance (Harmon & Julesz, 1973; Bindemann et al. 2013).

The experiments add to the existing knowledge of the familiarity advantage for accurate
face recognition. My study provides the first experimental investigation into the
familiarity advantage for distinguishing between true match and lookalike faces, finding
that familiarity does indeed aid this challenging task. I have shown that the graded
familiarity advantage extends to a more detailed breakdown of familiarity levels than had
been previously explored. This more detailed analysis was made possible by
acknowledging the idiosyncratic nature of familiarity – different viewers know different
faces to different degrees – and by allowing this insight to inform the design and analysis.

Too much pixelation destroys the graded familiarity advantage. Although an overall
familiarity advantage did carry through to the case of coarsely pixelated images (20x30),

83

this advantage was carried solely by the same identity pairs. This eventual breakdown
notwithstanding, the general robustness of the familiarity effect against declining image
quality suggests that familiar viewers are using information other than purely fine details
and precise configural information to support their make face-matching decisions.
Celebrities were used as the target faces in the experiments. It is therefore possible that
the results of the study would differ at the extremes of familiarity if photographs were
taken from people’s everyday encounters, rather than celebrities, to provide the target
faces. For example, performance may be better preserved if the target face was a family
member, or worse if the target face was someone who had never been seen before
(participants may have had some prior exposure to the celebrity faces even if they were
not aware of it).

As well as their theoretical interest, these findings may also have practical implications.
Forensic investigations regularly rely on face images as a means of evidence (Loftus &
Doyle, 1992). I found that the more familiar a viewer was with the target face, the better
was their ability to reject similar faces. It seems logical that in situations involving identity
fraud or poor quality images that a viewer who is of the highest available level of
familiarity with the target face would be best placed to judge the identity of the person
concerned – and that a little familiarity may be better than none.

In summary, increasing familiarity with a target face increases a viewer’s ability to
integrate different images of the same person and to distinguish images of different
people – even in the context of very similar faces, and poor quality ambient images. I
approach a limit to this familiarity advantage, where increased familiarity cannot fully
compensate for reduced image quality. In the next chapter, I consider how performance
on this difficult face-matching task might be improved.

84

There is also practical relevance to this experimental series. Performance was identified
to be significantly poorer as a result of image quality degradation, yet there are situations
when poor quality images are all that are available to aid an investigation. It is therefore
of interest to find ways of improving performance for the reduced quality images used in
these experiments. This will be investigated in Chapter 3.

85

Chapter 3 – Improving Performance
3.1 Chapter Summary
In this chapter I test several ways of improving the poor performance for pixelated images
seen in Chapter 2. Pixelated images of faces are often encountered in forensic
investigations when zooming in on digital images. Thus improving performance in this has
highly applied relevance. I show that blurring the pixelated images, and applying crowd
analysis can both improve performance on a pixelated matching task. Moreover,
performance benefits due to these were additive meaning that both could be used
together for even greater performance improvement. Finally, I found that superrecognisers outperformed control participants, even in the extremely challenging task of
imposter detection for poor quality images. This is the first time that super-recognisers’
performance has been shown to extend beyond good quality images of cooperative
stimuli.

3.2 Introduction
In Chapter 2, I showed that familiarity can help when dealing with pixelated images, but
this familiarity advantage was pushed to its limit when dealing with the coarsely pixelated
images in Experiment 3 (20x30 pixels). In forensic situations, the problem of identifying
pixelated faces is often encountered because this is the resulting image type from
zooming in on CCTV footage (Bindemann et al., 2013) or other digital images (Jenkins &
Kerr, 2013). In many such cases, finding a viewer who is familiar with the faces concerned
is not a viable option. For example, passport security officers or club bouncers may be
required to make identity judgments concerning many individuals in a very short period
of time. With poor image quality being a very real problem in applied environments, it
would be useful to find ways to improve human face recognition performance for
pixelated images across the familiarity continuum. In this chapter I will test several very
different methods for improving pixelated face recognition, specifically image
manipulations, data analysis and specialist viewer groups.

86

Most previous attempts to improve face recognition performance have revolved around
training, usually using good quality images. However, training approaches have met with
little success. In an early example, Woodhead, Baddeley & Simmonds (1979) evaluated a
longstanding three-day training course, which aimed to improve face recognition
performance of attendees. The course was deemed intensive, consisting of lectures,
demonstrations, discussion and practical work. Specific focus was placed on learning
isolated features, as the course founders believed this to be the key to successful face
recognition. Before this study was conducted, the success of the programme had not
been measured. To measure its effectiveness, attendees were tested on face-matching
and face memory tasks both before and after the three days of training. Their
performance on these tasks was compared with a control group who did not take part in
the training course. It was concluded that undergoing training did not significantly
improve performance in any of the test tasks, with both trainees and controls showing
similar mean hit rates between .6 and .9 depending on the task type. In one test training
actually led to a significant decrease in performance compared to controls. The authors
explained this by suggesting that attention to isolated features may impair face
recognition performance rather than improve performance in some instances (Woodhead
et al. 1979). This explanation for poorer performance after training is built on the
research of Winograd (1976) who found that when participants focused their attention on
one specific facial feature, their memory for the face was impaired.

Figure 3.1 Face shape classification examples provided by Towler et al. (2014).

Overall Woodhead et al. (1979) showed that attendance on a training course which
focused on attention to isolated features did not improve face recognition accuracy.

87

However, it is possible that training in other methods could help. Towler, White & Kemp
(2014) assessed the face shape training strategy, which focused on classifying the shape
of a face (e.g. oval, round, square) as the basis for successful face recognition (see Figure
3.1). Unfortunately this technique was also not successful in improving face recognition.
Participants of the course were no better at the GFMT after undergoing the face shape
training than they were before training. The face shape training strategy seems to be
fundamentally flawed, as a face does not take on a consistent shape and also, it is difficult
to classify faces by shape as there are not particularly clear distinctions between each
category e.g. what one person considers a round face another may consider to be oval.
This was found to be the case in the study, as different images of the same individual’s
face were frequently described as having different face shapes. During training
participants viewed a series of five different same identity photographs. Each identity was
judged as having the same face shape across all five photographs in only 7% of cases. It
was noted that the perceived shape of a face was not a diagnostic characteristic of
identity, hence explaining why face shape training does not improve face recognition. I
return to the issue of face shape in Chapter 4.

Some training studies have shifted away from identifying specific aspects of a face, to
more general strategies. White, Kemp, Jenkins & Burton (2014) showed slight but
significant improvements in face-matching performance when participants received
immediate feedback on their decisions. Here the authors were not concerned with how
the viewers made their decisions, but instead focused on notifying viewers on whether
identity judgments they made were right or wrong. This worked in the form of
participants’ receiving a correct or incorrect statement, immediately following their
answer submission for each trial of a face-matching task that they completed. The images
remained on the screen while the feedback was presented and improvement generalised
to new faces shown in the task, but it is not known how long lasting the benefit from the
feedback would be.

88

Taking a rather different approach, Dowsett & Burton (2014) also report some positive
effects of working in pairs when making matching identity judgments. Individuals were
tested for face-matching performance, and then tested as part of a pair, and finally tested
individually again. In the pair judgment condition, the identity decision was made
collaboratively after discussion. Working in a pair, lifted performance to the level of the
higher individually performing member. Interestingly, the effect of pair working carried
over to improve the performance, especially of the weaker pair member, at the later
individual testing. This study, along with the work of White et al. (2014), provide evidence
for feedback as an important self-regulator of performance accuracy, which may improve
face recognition in some situations. Although performance improvements were
statistically significant, they were numerically small in both studies.

Although training courses have been found not to improve face recognition performance,
and recent lab based studies showed only small improvements, it remains a widely held
belief that those who hold jobs that rely on the ability to accurately identify faces, will be
better at face recognition than people whose jobs do not rely on this ability.

In reality, it seems that highly trained officials are no better than untrained and
inexperienced others at matching faces. Burton et al. (1999) reported that police officers
performed at the same level of accuracy as undergraduate students in a task that
involved matching poor quality CCTV footage images to comparison face images. More
recently, passport officers have been tested on the GFMT. Passport officers’ performed
with 79.2% accuracy on the task, whereas mean performance for the general population
control group was 81.3% (Figure 3.2). There was no significant difference between these
performance scores. It could be argued that the GFMT task does not mimic the identity
matching task that passport officers perform, which involves comparison of a photograph
to a physically present face rather than to another static image. White et al. (2014a)
addressed this by testing passport officers on a task that directly mimicked the passport
control scenario of matching an image to a physically present face, and also on image-toimage matching performance. Passport officers again performed no better than a control

89

group of undergraduate students. Relating back to the work discussed above on training,
no relationship was found between number of years on the job and performance
accuracy. These findings make sense with reference to previous research – if training
holds little benefit elsewhere, there is no reason, other than perhaps increased
motivation, why similar training would have improved performance for those people
whose jobs rely on high face recognition accuracy when they have not helped before.

Figure 3.2 Graph from White et al. (2014) passport officer paper showing the officers’ performance
accuracy on the GFMT alongside their employment duration. Some police officers performed very highly on
the GFMT, these high scores can be found at both ends of the employment duration axis.

Taken together, the previous research on improving face recognition performance
suggests that training is unlikely to improve face recognition performance on challenging
pixelated images. In this chapter I will attempt to improve face recognition performance
using three distinct approaches, none of which rely on training. There are three methods
that I will test for improving face recognition performance. First I will investigate whether
image manipulations, in the form of blurring pixelated images, can improve performance.
Second I will examine whether crowd effects, which is a data analysis technique, can be
applied to data that has already been collected to improve performance. Third, I will test
whether super-recognisers can be relied upon to make more accurate identity judgments
than controls for challenging images. Each of the approaches I use, and past research
involving these techniques will be described and discussed in detail within the body of
this chapter.

90

3.3 Experiment 4: Blurring Pixelated Images
It was evident in Chapter 2 that although an advantage of extreme familiarity survived for
face recognition performance when dealing with coarsely pixelated image pairs, overall
performance was poor compared to the prior better quality image pair experiments.
Performance was at around chance level for faces that were unfamiliar, and poor for the
extremely familiar faces relative to performance for this familiarity level in Experiments 1
and 2 of Chapter 2.

It has been noted by past researchers that ability to recognise a pixelated face can be
improved by blurring the image. It may sound somewhat counterintuitive to blur a
pixelated image, because blurring removes information from the image. Harmon & Julesz
(1973) explain that when an image is pixelated, each square is a result of the average
density of the pixels that makes up this area in the original image. There is more of a
difference in amplitude between two adjacent pixels than there may have been between
two of the pixels in the original image. This difference in amplitude introduces high
frequency noise at the pixel edges, making it difficult to extract useful information.
Configural information becomes less precise and featural information less detailed.
However, when a pixelated image is blurred, high frequency noise is removed and
identity is easier to recover (Harmon & Julesz, 1973; Morrone, et al. 1983). Blurring is
essentially the same as low pass filtering, in that both processes filter out high spatial
frequency information.

Here I examine whether blurring the pixelated images used in Chapter 2, Experiment 3
will improve performance despite the extreme similarity of the face images in each pair. If
blurring enhances performance on the task then this technique could be used in applied
settings to aid facial identifications in forensic investigations.

91

I predict that removing high frequency image noise through blurring the pixelated images
will improve accuracy as compared against the results of Experiment 3.

Method
Participants
A group of 30 undergraduate students (M= 6, mean age = 19.7) at the University of York
(who had not taken part in any of the previous lookalike tasks involved in this series of
experiments) volunteered as participants for this study.

Design and Stimuli
As in the previous experiments, this experiment adopted a within subjects design. The
variable familiarity was examined at three levels in order to keep consistency with the
design of Experiment 3. This allowed me to perform a between experiments comparison
of performance for the different image types (pixelated and blurred), comparing the
results of this experiment from those from Experiment 3.

Face-matching Task
The stimuli for this experiment were modified versions of the face images from
Experiment 3. To create the stimuli versions necessary for this new experiment, I took the
pixelated images used as the stimuli in Experiment 3 and applied a blurring technique to
these images using Adobe Photoshop (CS6). The 20x30 pixel face images were blurred at
a radius of 3.8 pixels using Photoshop’s Gaussian blur function (see Figure 15 for a side by
side comparison of the coarsely pixelated and blurred-pixelated stimuli). This blurring
level was determined via pilot testing in which two raters assessed changing pixelation
levels on a sliding scale and decided by eye on a level that they believed made the image
easier to identify.

92

Figure 3.3 Identical images of Al Gore (left) and Gary Barlow (right) shown as they were presented in each
experimental condition. The image on the left for each identity, shows the coarsely pixelated image as
presented in Experiment 3. The images on the right, show the image on the left of it, after undergoing
blurring, and as presented in Experiment 4.

Procedure
The procedure was the same as for Experiment 3 except that the images viewed were a
blurred version of the coarsely pixelated images. As in the prior experiments, participants
ranked each of the celebrity faces in order of familiarity using the familiarity scale (see
back to procedure section in Experiment 1 for a detailed description of the face-matching
task and familiarity rating scale used for all experiments in this series).

Results
There was a significant main effect of Familiarity on face-matching performance for the
blurred version of the task [F(2,58) = 6.07, p<.01, η p2 = .17]. Participants performed with
lowest accuracy for familiarity band 1 (M = 62%, SE = 2.15), with accuracy increasing as
familiarity increased for band 2 (M = 65.5%, SE = 2.5) and band 3 (M = 72.3%, SE = 1.81).
Performance for familiarity band 3 was significantly greater than for band 1 and band 2,
no other differences were significant.

93

In Experiment 3, the familiarity advantage was found only for same identity trials. In order
to check whether blurring recovered the graded familiarity advantage for each type,
performance accuracy for the blurred pixelated faces was analysed according to same and
different trials using a mixed ANOVA.

Figure 3.4 Graph showing performance accuracy on the blurred pixelated task split by same and different
person trials. Error bars show the standard error of the mean.

For the Same and Different identity trial breakdown there was a significant main effect of
Familiarity [F(2, 58) = 6.02, p<.01, η p2 = .17] but no main effect of Trial Type [F(1,29) = .96,
p=.33, η p2 = .03] and no interaction between Familiarity and Trial Type [F(2,58) = .13, p =
.87, η p2 = .01]. This shows that for the blurred version of the faces, familiarity improved
performance on same person trials and on different person trials (see Figure 3.4).

94

Between Experiment Analysis
To find out whether performance was better for pixelated and blurred faces than for
pixelated faces, a between experiments analysis was conducted using a 2 way mixed
ANOVA. Factors for the ANOVA were familiarity (within subject factor) and image type,
pixelated or pixelated and blurred (between subjects).

Figure 3.5 Percentage of correct responses in face-matching task for each familiarity band (low familiarity,
mid familiarity, high familiarity), for Experiments 3 (black line) & 4 (blue line). Error bars show standard
error of the mean.

Overall, performance accuracy was significantly higher when pixelated images were
blurred (M = 66, SE = 1.36, CI = 64. -70) than performance had been in the Experiment 3
which consisted of the coarsely pixelated images (M = 59, SE = 1.36, CI = 56 -62), [F(1,58)
= 16.75, p<.001, η p2 = .22], see Figure 3.4.

95

Additionally, a main effect of Familiarity was evident from the results [F(2,116) = 12.53,
p>.001, η p2 = .18]. Participants were better at recognizing the faces they were most
familiar with (M = 69, SE = 1.27, CI = 66-71) than the faces that were of Mid (M = 60, SE =
1.89, CI = 56-64) or Low familiarity (M = 59, SE = 1.34, CI = 57-62) to them. No interaction
was observed between Familiarity and Experiment [F(2,58)=1.33, p=.27, η p2 =.02].

To keep analysis consistent with that of previous experiments, a three way ANOVA was
conducted for Familiarity (within-subjects, 3 levels – Low familiarity, Mid familiarity, High
familiarity), Trial Type (within-subjects, 2 levels – same, different) and Experiment
(between subject, 2 levels – Experiment 3, Experiment 4). The ANOVA revealed a
significant main effect of Familiarity [F(2,116) = 11.2, p<.001, η p2 = .16]. As expected from
Figure 2.11 and Figure 3.5 there was significant three way interaction between
Familiarity, Trial Type and Experiment [F(2,116) = 3.53, p<.05, η p2 = .06] confirming that
Familiarity affected performance according to Trial Type in different ways across the
experiments. No other results were significant.

Discussion
Blurring the pixelated images had a significant positive effect on performance accuracy.
Overall, performance was 12% better when participants viewed blurred versions of the
pixelated images (Experiment 4) compared to when participants viewed the pixelated
image (Experiment 3).

Previously, blurring a pixelated image had been found to aid identity recognition of a
pixelated face (Harmon & Julesz, 1973; Morrone, et al., 1983). My research found that
this blurring advantage extended to matching involving similar images. Blurring the image
to a greater or lesser degree may have led to more of an improvement, but I was

96

primarily concerned on whether or not any blurring would improve performance, which it
did.

My experiment also leads to theoretical implications, helping to answer whether identity
information is carried in high or low spatial frequency information in a face and more
specifically, how this information plays out in the identifying similar faces. When a
pixelated image is blurred, the high spatial frequency information in an image is removed.
It was previously reported that faces could be identified from blurred pixelated images,
which suggests that it is the low spatial frequency information which is important for
identity (Harmon & Julesz, 1973; Morrone et al. 1983). My task involved very similar
images therefore any differences between the identities was subtle. As blurring improved
performance for these images, one interpretation of my findings is that some of the
subtle differences between identities were held in low spatial frequencies.

The findings of this study are of great practical relevance. Pixelated images are often the
image type which police have to deal with. It has been shown by previous studies that
performance on unfamiliar face-matching tasks is extremely poor, especially when the
images are of poor quality. Experiments 1, 2 and 3 reiterated this, with the extreme and
more challenging case of imposter detection. I found that performance was particularly
poor in the imposter detection task for the unfamiliar faces, and performance
deteriorated as a function of image quality. Blurring the pixelated images is a simple
image manipulation that police could use to aid the likelihood of correctly distinguishing
and detecting same and similar faces.

3.4 Experiment 5: Crowd Analysis
In some situations data on identity tasks has already been collected before the
challenging nature of the task due to pixelation has been addressed through image
manipulation. It is possible that more accurate results of identity judgments can be

97

obtained from the existing data, through reanalysing the data already collected but in a
different way. We have seen already that performance is poor, even for familiar viewers,
when the images are coarsely pixelated. Blurring the pixelated images, as above, did
improve face recognition performance, however accuracy remained lower than it did for
good quality images. A technique called crowd analysis, may improve performance
accuracy, and could even help in the situation when no viewers are extremely familiar
with the faces whose identities are in question.

Crowd analysis is based on the idea that a group performs better than an individual
(Galton, 1907; Krause et al., 2011; White, Burton, Kemp & Jenkins, 2013). Rather than
relying on a single person’s decision, the majority vote can be obtained by pooling these
individuals’ results to find the group average response. The most commonly given answer
is taken as the group answer – simple majority rule. Crowd effects are thus created by
pulling together the mean result of a group of people. This mean answer is then taken as
the new response and compared against the correct answer.

The power of the crowd over an individual has been acclaimed for quite some time;
Aristotle addressed the wisdom of the crowds in his book Politics (Aristotle, published
1920). Later, statistician Sir Francis Galton (1907) demonstrated that a crowd could
outperform the individual by calculating the median estimated weight judgment of an ox
in a guess the weight of an ox competition at a fair. He found that the median score of a
crowd of 800 people provided an answer that was within 1% of the true weight of the ox.
Since Aristotle and Galton, the concept of wisdom, or power of crowds has been referred
to by many different names. In biological terms, this phenomenon is commonly referred
to as swarm intelligence, and refers to the superior ability of groups in solving cognitive
problems over the individuals who make up that group. The benefit of swarm intelligence
has been more recently replicated through experiments involving guessing the number of
marbles or the number of sweets in a jar (Krause et al. 2011; King et al. 2012). An
interesting finding in these studies was that groups of low performers could outperform
individual high performers. Krause et al. 2011 highlight that swarm intelligence is

98

beneficial in providing a more accurate answer to a sweet in a jar question, but less
beneficial when dealing with questions that require expertise of knowledge – this was
tested with regards to predicting coin toss odds in relation to the odds of winning the
lottery. For the coin toss and lottery question, groups would only outperform experts if
the group size was larger than 40 people. Group sizes varied between 2 and 80, with
larger group sizes providing more accurate results.

King et al. 2012 also found that swarm intelligence improved accuracy of sweet guessing
judgments, but then adapted the experiment to find out whether providing the individual
guessers with additional information (the average guess taken by the people before
them, the guess of a randomly chosen previous person, the guess of the person before
them or the best guess which had been made before theirs) would sway their judgment.
With no additional information, the crowd far outperformed the individuals. Knowledge
of any of the previous guesses, except the case of the best previous guess, reduced the
effect of swarm intelligence. Access to the best guess led to more accurate results than in
each of the other conditions at both and individual and small group analysis level. This
suggests that crowd analysis does help, in all situations, but the added information of an
expert’s guess could help improve accuracy further.

White et al. (2013) were the first to have used the wisdom of the crowd theory to address
the problem of face identity judgments, and refer to the technique they use as crowd
analysis. The principle remains the same; the most common result of the group is taken
as the answer (majority vote) instead of averaging responses at an individual response.
White and colleagues (2013) report that performance accuracy on the GFMT can be
improved by analysing the results of the GFMT (Burton et al. 2010) for crowd sizes of 2, 4,
8, 16, 32 and 64 subjects, rather than looking at the mean overall response at an
individual level as previously reported (see Figure 2.3). It is important to note that the
authors are dealing with the exact data collected from Burton and colleagues (2010)
study, where mean performance for individual accuracy was 89.9%, it is only the type of
analysis which has changed. Crowd analysis, for groups as small as four people,

99

outperformed mean individual accuracy. Group size was also important, with larger
groups providing more accurate results of identity matching than when individual results
were pooled across smaller groups. The largest group size of 64 led to performance
accuracy levels of 99.2%. Further still, crowds outperformed the highest performing
individual when the data was aggregated over a group size of eight or more (White et al.
2013).

Figure 3.6 Mean performance on items of the GFMT performance according to different crowd sizes (White
et al. 2013). Graph shows performance accuracy broken down by trial type, with results analysed for crowd
sizes of 1, 2, 4, 8, 16, 32 and 64.

This study by White et al. (2013) showed for the first time that crowd analysis can
improve performance on face identity judgments. The training techniques discussed in
the introduction of this chapter focussed on improving performance at the individual
level but this led to no or little success. Data analysis techniques may actually allow us to
improve the accuracy of identity decisions without the need for training, and even to
improve accuracy retrospectively, on tasks for which data has already been collected. This

100

is useful for new studies also, as data collection methods would not need to change in any
way. The crowd analysis technique has proved useful in tasks involving good quality
images (e.g. White et al. 2013), but has not yet been applied to data from tasks involving
more challenging images.

I will test whether performance can also be improved through crowd analysis for my
pixelated lookalike task, which includes very similar and poor quality images. It is not yet
known whether the crowd advantage can extend to help in this very challenging case of
identity judgment. This is particularly interesting to investigate, as a more challenging task
will allow a better understanding of the advantage that crowd analysis can hold. In the
GFMT, crowd analysis could only improve individual performance by a maximum of 10%,
which brought performance to ceiling level (see Figure 2.3). It is unknown whether crowd
analysis could improve performance by more than 10%. Performing crowd analysis on my
more challenging task will allow this to be investigated, as the lower baseline
performance level in this tasks allows far more room for improvement than in the GFMT.

Based on the success of crowd analysis in previous situations, I predict that performance
accuracy on the pixelated lookalike task will improve as crowd size increases.

Method
Data
The dataset used in this study was the raw data from Experiment 3, the coarsely pixelated
lookalike task, (N = 30, M = 8, mean age = 20.2).

101

Analysis
To carry out crowd analysis I calculated the most frequently given response (same or
different) for each item of the coarsely pixelated face-matching task, Experiment 3, across
different crowd sizes (i.e. subgroups of participants) and then used this response as the
answer for that item. Crowd responses were calculated for all items tested to determine
overall percentage accuracy. Results were calculated across crowd sizes of 1, 3, 5 and 15,
these crowd sizes were selected as denominators of 30, which allowed the creation of
equal group sizes given that there were 30 participants in the study.

Results & Discussion
To find out whether mean performance accuracy increased with increasing crowd size,
individual performance (crowd size 1) was compared with performance of crowds of 3, 5
and 15 (see Figure 2.4).

102

Figure 3.7 Graph showing the mean accuracy score for crowd sizes of 1, 3, 5 and 15 for the coarsely
pixelated lookalike task, Experiment 3. Error bars show the standard error of the mean.

Individuals (M = 60.11 % correct) were outperformed by each of the crowds, with higher
performance for larger crowds (crowd size 3, M = 64.16% correct; crowd size 5, M = 65 %
correct; crowd size 15, M = 67.5 % correct). The crowds were made up of the same
people who contributed to the individual performance analysis, at no point was any
individual becoming good at the task - crowd scores were calculated after independent
decisions had already been made. These findings show that crowd analysis did improve
performance levels, with increasing crowd size leading to better performance.

In keeping with previous analysis, results were broken down into same and different
trials. Increasing crowd size increased accuracy for Same identity trials but not for

103

Different identity trials (See Table 3.1). These results mirror those of trial type reported in
Experiment 3.

1(SE)

3(SE)

5(SE)

15(SE)

Same

62.2 (.55)

66.7(2.85)

70(3.44)

73.3(3.33)

Different

58(.59)

61.7(3.8)

60(2.11)

57(5)

Table 3.1 Crowd analysis results broken down by trial types (same identity, different identity).

Overall, performance improved through crowd analysis, but did not exceed a 10%
increase in performance, which would have been possible given the task difficulty. The
coarsely pixelated celebrity lookalike task was a very difficult task, which is shown
through the highest group performance being just 70% accuracy. It is possible that
performance could be improved even more if I apply the crowd analysis technique to the
results from the blurred pixelated faces, Experiment 4. I have shown that blurring
improved performance on the lookalike task, possible that crowd analysis on the slightly
easier, blurred, version may lead to even better performance than either crowd analysis
applied to the coarsely pixelated faces, or blurring the faces, alone.

Combining Improvement Techniques (Experiment 5b)
Blurring pixelated images (Experiment 4) and performing crowd analysis to pixelated
images (Experiment 5) have both resulted in improvements in performance accuracy
compared to performance on the coarsely pixelated celebrity lookalike task (Experiment
3). Both blurring and crowd analysis had been proved successful in improving facematching accuracy in past studies, yet no work has attempted to combine these
techniques, i.e. perform crowd analysis on the blurred pixelated image data.

I predict that crowd analysis will have a similar effect on the blurred pixelated data as it
did on the coarsely pixelated data, therefore increasing crowd size for the blurred
pixelation data will lead to more accurate results.

104

Method
Data
The dataset used in this study was the raw data from Experiment 4, the coarsely pixelated
lookalike task, (N = 30, M= 6, mean age = 19.7).

Analysis
Crowd analysis was carried out on data from the blurred coarsely pixelated celebrity
lookalike task, Experiment 4. The analysis procedure was carried out in the exact same
way as in Experiment 4 above.

Results & Discussion

To assess the effect of crowd analysis on the blurred pixelated images, crowd analysis was
calculated and results compared for increasing crowds. Crowd analysis improved upon
individual performance and improvements increased as crowd size increased. The crowd
size of 1 (M = 67.28% correct) was outperformed by crowds of 3 (M = 72% correct), 5 (M
= 75% correct) and 15 (M = 80% correct). These results are illustrated alongside the
results of Experiment 5 in Figure 3.8.

Crowd analysis was broken down by trial type for the blurred images (Table 3.2). When
the images were blurred, crowd analysis improved performance for both same and
different trials.

105

1(SE)

3(SE)

5(SE)

15(SE)

Same

64.3(.67)

68(2.29)

70.56(2.5)

73.3(3.33)

Different

70.2(.46)

75(1.87)

80(2.11)

87(0)

Table 3.2 Crowd analysis results broken down by trial types (same identity, different identity).

Next, results of blurred crowd analysis were compared with the results of the coarsely
pixelated experiment crowd analysis.

Figure 3.8 Graph showing the mean accuracy score for crowd sizes of 1, 3, 5 and 15 for blurred version of
coarsely pixelated lookalike task (blue line) and the coarsely pixelated lookalike task (black line). Error bars
show the standard error of the mean.

Performance gain due to blurring (the difference between the black and blue lines in
Figure 2.5) and performance gain due to crowd effects (represented by the slope of the

106

lines in Figure 2.4) are additive. This pattern suggests that the two gains have
independent causes (as demonstrated by the results of Experiments 3 & 4 broken down
by Trial Type). It also implies that both techniques can be used in combination to secure
the benefits of both. Indeed, using both techniques together leads to a large (20%)
accuracy gain in performance, from 60% accuracy for a crowd size of 1 in the coarsely
pixelated condition, to 80% accuracy for a crowd size of 15 in the pixelated and blurred
condition. The significance of this gain in performance is apparent when compared with
the gain of previous techniques such as training, which have improved performance
accuracy on previous face-matching tasks by around 5%. Overall, crowd analysis and
blurring has taken mean performance of individual viewers from 60%, which was low
compared to both unfamiliar viewers on easier (unedited) no pixelation versions of the
task, and compared to the performance of familiar viewers on any of the tasks, up to the
level of mean performance accuracy level of viewers who were fairly familiar with the
faces concerned in the easiest, no pixelated version of the task, Experiment 1. This
highlights that blurring and crowd analysis applied in combination greatly improving
performance for difficult images.

My results are important with regards to methods for face improving matching accuracy.
In past research, large increases in performance have been linked to the use of familiar
viewers in a task (e.g. Burton et al. 1999). The success of blurring and crowd analysis
techniques does not rely on familiarity with the faces in the matching task. Crowd
analysis, combined with blurring where appropriate, could thus aid judgments of very
difficult identity decisions and provide a solution to improving face recognition
performance that does not rely on familiarity with a face.

3.5 Experiment 6: Observer Factors
Unfamiliar face recognition ability is generally poor. We have seen this both in past
research (Burton et al. 1999; Bruce et al. 2001; Kemp et al. 1997) and in the results of the
thesis up to this point. There is however large variation at an individual level in face

107

recognition performance – some people consistently perform with very high levels of
accuracy, some with very poor levels of accuracy, and others at intermediate levels. This
was seen in the study by White and colleagues (2014), which tested face-matching
performance for passport officers. It was reported that although training and years of
experience made no difference to performance, there was a huge amount of variation in
performance amongst the passport officers themselves (White et al. 2014a). Face
recognition is therefore generally thought of as an ability that lies on a spectrum. At the
extreme lowest end of the face recognition ability spectrum are people with congenital
prosopagnosia (Duchaine, 2011). Congenital prosopagnosics have no identified brain
deficits, yet experience clinical level of difficulty with recognising faces (Berham & Avidan,
2005; Duchaine & Nakayama, 2006). At the other end of the spectrum are people who
consistently perform with exceptionally high accuracy. These people have been termed
‘super-recognisers’ (Russell, Duchaine & Nakayama, 2009). Super-recognisers may
provide a solution to the face recognition problem in practical settings. In theory, superrecognisers could help to eliminate the number of face-matching errors made, in
situations where face-matching is of high importance. If super-recognisers truly make
fewer errors than others, which is what these past studies have shown (e.g. Russell et al.
2009), then employers should consider testing candidate’s face recognition ability, and
making super-recogniser performance a requirement for the job roles for which accurate
face identification carries high importance.

Members of a highly specialised expert forensic group in the USA have however been
found to outperform controls on tasks involving face recognition. The ability of these
group members has only recently been identified. Indeed, White et al. (2015b) showed
that forensic examiners in the USA performed more accurately in three face-matching
tasks than control groups of trained experts in biometric systems (referred to in the study
as ‘controls’) and also undergraduate students (referred to as ‘students’). The trained
experts in biometric systems were highly motivated with the task, suggesting that
differences in performance were not due to differences in motivation. The face-matching
tasks used were the GFMT and two new tests, one which was specially designed to be
challenging to both computers and humans [EFCT] and the other made of stimuli which

108

contained cues that would be of use to humans only [PICT]. It is unclear so far whether
people gravitate towards the job of forensic examiner as a result of innate face
recognition ability, or whether the extremely specialised training helps these forensic
officers to perform highly on the face recognition tasks. As highlighted earlier in this
chapter, training has a poor track record for improving face identification performance
(Woodhead et al., 1979; Towler et al., 2014). The lack of success of previously evaluated
training methods would incline one to believe that an innate ability for face recognition is
key. It seems from the findings that the forensic examiners were using different
techniques to the comparison groups to make their identity judgment. This could be a
result of innate ability or training. It would be interesting to find out exactly what the
training for these forensic examiners involved, and this knowledge may help to answer
whether the examiners’ face recognition ability is innate or learnt.

Additionally, further testing and analysis of the data collected from White and colleagues’
(2014) study on passport officers’ face-matching performance, revealed that a subset of
the passport officers were identified as particularly high performers. These highly
performing officers were in a specific branch of the passport office, in a job role where
their ability to successfully match true match faces and identify false matches was of
paramount importance. Thus, high face recognition ability was even more important for
their job role than for the roles held by other passport officers tested (White et al. 2015a).
It remains unknown whether the officers came to their specific role due to being naturally
best suited, or if they learnt the skills required on the job. This has not been investigated
directly to date. As years of training was recorded, and found to have no effect on
performance accuracy, it may be more likely that innate ability guided these passport
officers into their specialised job, where face accurate face recognition is extremely
important. As in the case of the US forensic examiners, more information about the
training programmes and ideally also new data that specifically tracks only the members
of these highly specialised expert groups over the course of training and job experience,
could help to answer whether ability has been innate, learnt or perhaps even a
combination of both.

109

The performance of super-recognisers had until recently not been tested and directly
compared to controls by standardised tests. Instead, super-recognisers were identified
from large groups of people tested on standardised tests, as those with the very highest
scores. Bobak, Dowsett & Bate (2016) specifically tested the performance of self
identified super-recognisers to find out whether they were better than a student control
group at two standardised tests - the GFMT and a face-matching task involving models.
The models task is challenging, as the same models can look very different across
different images. All of the model faces were unfamiliar to the viewers, therefore this
study tested purely unfamiliar face recognition performance of super-recognisers. The
authors reported that their group of 7 self-identified super-recognisers were better than
a student control group at the two tasks. This finding suggests that people who have
exceptional face-recognition ability are in fact aware of this ability, as their superior
performance to others was evident when tested on standardised face identification tasks.

Super-Recognisers within the MET Police
The evidence provided overwhelmingly supports the use of super-recognisers as a
possible solution to the face identity problem. It therefore seems sensible that face
identity professionals consider employing super-recognisers. This advice has been taken
by The Metropolitan Police who have recently established a super recogniser team within
their force. This has been created by internally recruiting officers who have particular
interest in face recognition and performed well in their undisclosed face recognition test.
These officers regularly try to identify faces from pixelated CCTV photographs. During the
establishment of the super-recogniser team I was given the opportunity to test four of
these MET police super-recognisers in their normal working environment, to see whether
they outperformed our undergraduates on the lookalike task. The lookalike test is
particularly relevant for measuring face recognition ability of the police super-recognisers
as it tests performance of faces of differing levels of familiarity – from unfamiliar to
extremely familiar. The police have to deal with cases from both levels of this spectrum
depending on whether the individual involved in the crime is a known repeat offender or
committing their first crime. Due to time constraints I could test performance on just one

110

version of the test. I wanted to test performance on one of the pixelated lookalike tasks,
as Chapter 1 demonstrated that even familiar undergraduate viewers do not score with
perfect performance on these tasks. Using pixelated images would therefore avoid ceiling
effect for matching trials. Pixelated images are also an image type that the police have to
deal with as part of their investigations, meaning that the task holds practical relevance.
The mid pixelation level (30x45 pixels) was chosen so that predicted performance would
be at a level expected to be above chance as based on previous findings (the higher
pixelation level reduced performance to chance for all but extremely familiar faces in our
experiment on undergraduate students), this also guaranteed that I had an already
existing comparison group whose performance was above floor (>50% accuracy).

Participants
Participants were 4 super-recognisers (M=4, mean age = 40) from the Metropolitan police
force super-recogniser team, New Scotland Yard Central Forensic Image Unit, London.
Performance was compared with that of our 30 undergraduate students tested in
Experiment 2 (M = 11 male, mean age = 19.7).

Design & Stimuli
The stimuli were images of the 30 celebrities and lookalikes for these celebrities used in
Experiment 2. As the comparison viewers were generally younger than the superrecognisers, I took care to ensure that any differences in performance could not arise
through some celebrities being more familiar to one group or the other.

As in Experiment 2, all images were presented in low resolution (30 pixels wide x 45 pixels
high). This resampling also served the purpose of reducing matching images to the level
that we expect to avoid ceiling effects in the matching trials. Examples of the stimuli are
shown in Figure 3.9.

111

Figure 3.9 Example trials from the PLT. Images on the left show different identities (with the imposter face
on the right). Images on the right show the same identity.

Procedure
The procedure was similar to that used in Experiment 2. The police super-recognisers
viewed the images in a printed booklet rather than on a computer screen. Superrecognisers were presented with a printed booklet containing 60 trials, half of which
showed the same identity and half of which showed different identities. They were asked
to make a same/different judgement for each trial. Following the face-matching task
participants used a numerical scale to indicate their level of familiarity with each celebrity
whose face had been viewed in the task (from 1 [completely unfamiliar] to 10 [extremely
familiar]).

Police super-recognisers were also tested on the GFMT and the models face-matching
task (MFMT) which were designed and administered by other researchers from the
University of York Facelab.

112

Results

Figure 3.10 Performance of police super-recognisers and comparison viewers. Performance of super-
recognisers (SR1–4; black) and comparison viewers (white) on three different tests of face recognition—the
GFMT (left column), the MFMT (middle column), and the PLT (right column). Vertical lines indicate the
range of scores for comparison groups, the deleted portion of the line shows the standard deviation, and
the horizontal notch shows the mean. In all three tasks, chance performance is 50%.

In Experiment 2, results were broken down according to 5 levels of familiarity, with
results analysed according to these familiarity quintiles. Due to the small number of
super-recognisers that I was able to test, this time I broke familiarity down into two
groups (highly familiar [80-100 on familiarity scale] and less familiar faces [all other
ratings]) to ensure sufficient numbers of faces were placed in each familiarity bin.

Analysis was broken down according to the two familiarity bands, with the first tests
being performed on the subset of faces that viewers rated as highly familiar. Interestingly,
super-recognisers gave high familiarity ratings (80–100) to a very high proportion of faces
compared with controls (70% of faces for super-recognisers; 37% of faces for controls).

113

The police super- recognisers consistently performed with far greater accuracy than
student participants. Overall, for highly familiar faces the controls scored just 76%
correct, whereas the police super-recognisers scored with 93% accuracy in the lookalike
test. All super-recognisers performed much better than the control mean, with one
super-recogniser performing perfectly. This performance is shown in Figure 3.10 (for full
results of the GFMT and MFMT see Robertson, Noyes, Dowsett, Jenkins & Burton, 2016).

To better understand the super-recogniser advantage, I conducted two further analyses.
First, I compared accuracy of super-recognisers and controls on faces that they rated as
less familiar (0–7 on the 10-point scale; i.e. those not included in the above analysis).
Police super-recognisers outperformed controls on these faces too (76% accuracy for
super-recognisers; 66% accuracy for controls), implying that their performance advantage
holds across the whole familiarity continuum. Second, I analysed the control participants’
data for an association between i) the proportion of faces that were given high familiarity
ratings and ii) the level of accuracy on those highly familiar faces. A significant positive
correlation was found between these two measures [r(28) = 0.39, p < .05], such that the
highest performing controls were qualitatively similar to the super-recognisers. These
analyses support the original comparison, i.e. the super-recognisers are performing at
well above the levels of controls, even when group differences in famous face familiarity
are taken into account.

Result breakdown by trial type demonstrated that SRs outperformed controls at both
same and different identity pairs, for both the low and high familiarity groups (See Table
3.3).

114

% Accuracy

% Accuracy

Low Familiarity (SE)

High Familiarity (SE)

Overall

75.72 (8.52)

93 (5.12)

Same

83.65 (11.78)

95 (2.84)

Different

67.79 (19.09)

90 (3.14)

Controls

SRs

Overall

66.4 (1.73)

76.36 (2.81)

Same

68.98 (2.17)

81.34 (3.61)

Different

64.10 (2.97)

70.18 (3.54)

Table 3.3 Performance accuracy broken down by viewer group and trial type.

Crowd Analysis on Super-Recogniser Data
Earlier in this chapter I demonstrated that crowd analysis improved performance
accuracy on the coarsely pixelated face matching task. I also confirmed that combining
methods of improvements had additive benefits on performance accuracy. To continue
the investigation of the effect combining improvement methods, I performed crowd
analysis on the super-recogniser data. This analysis incorporated all items, regardless of
familiarity with the item (as in the crowd analysis earlier, Experiment 5). Mean
performance accuracy for crowd size of 1 (individual analysis) was 80%. Crowd analysis
improved performance, with mean performance accuracy levels of 88% for a crowd size
of 4 (see Table 3.5). This performance cannot be directly compared to the crowd analysis
in Experiment 5, as the police super-recognisers completed the mid level of pixelation
task rather than the coarsely pixelated version. Instead new analysis of Experiment 2 (the
mid pixelation level face matching task) was conducted using crowd analysis. Crowd
results for the mid pixelation group, Experiment 2, are shown in Table 3.4. Increasing
crowd size increased performance accuracy, from 62% for a crowd size of 1, to 83% for a

115

crowd size of 15. These results highlight the exceptional face recognition ability of the
super-recognisers, as a crowd size of just 4 super-recognisers outperformed the largest
crowd size of 15 controls (Table 3.5).

Crowd Size

1(SE)

3(SE)

5(SE)

15(SE)

Same

66.33 (2.14)

81.33(3.56)

83.88 (3.26)

88.3(5)

Different

58.77(2.1)

69 (5.24)

72.2(6.81)

78.3(1.67)

Table 3.4 Crowd analysis results broken down by trial types (same identity, different identity) for the mid
pixelation (control) Experiment 2.

Crowd Size SRs

1(SE)

4

Same

86.66 (3.85)

100

Different

74.12(7.67)

76.67

Table 3.5 Crowd analysis results broken down by trial types (same identity, different identity) for the SRs.

Discussion
This is the first time that police super-recognisers’ performance has been assessed using
standardised tests. Superior performance of super-recognisers in the pixelated lookalike
task suggests two important conclusions. First, the testing systems used by the police to
recruit super-recognisers are indeed successful in selecting people with exceptional face
recognition ability. Second, these super-recognisers’ ability is not limited to good quality
images; super-recognisers are superior to others in face recognition ability even in
extremely image conditions involving extreme similarity between target faces and foils.
The performance of super-recognisers can be improved even further through crowd
analysis. The largest crowd size of 4 super-recognisers was relatively small, but
outperformed even the largest crowd size of 30 control participants.

116

3.6 General Discussion
I have shown in this chapter that the poor performance found for the highly challenging
pixelated stimuli images can be improved upon through use of a variety of techniques.
Performance accuracy was improved by blurring pixelated images, combining results
using the crowd analysis technique, and using the results of super-recognisers. Blurring
and crowd analysis, and also super-recognisers and crowd analysis can be used in
combination for additional benefit.

Each of the techniques that I investigated had shown promise in improving performance
in previous tasks of face recognition. However, none of the techniques had been applied
to such a challenging identity matching scenario. Blurring had been studied in terms of
face recognition rather than identity matching (Harmon & Julesz, 1973). The only study of
face recognition to incorporate the crowd analysis technique analysed results from a
cooperative face-matching task involving good quality images (the GFMT). Superrecogniser performance has also been studied in terms of good quality images, for
unfamiliar faces only (Bobak et al., 2016). Past studies have also looked at each of these
techniques exclusively for unfamiliar faces. My findings support the body of existing
research, adding that the benefits of all of these techniques extend to very challenging
pixelated image conditions regardless of familiarity with a face. I also examined the effect
of combining techniques in the combination of blurring combined with crowd analysis,
and super-recognisers combined with crowd analysis. As far as I know, these different
techniques have never before been applied in combination. Combining the techniques led
to even further improvements than using any of the techniques in isolation. This suggests
that each of the techniques improve performance in a way that is different from each of
the other methods. Blurring reduced high spatial frequency noise from a pixelated image,
crowd analysis relies on the performance of the crowd being more accurate than the
individual and super-recognisers perform highly, but the methods that super-recognisers
use are unknown and may link to innate ability. As each of these methods aid the face

117

recognition problem in a different way, when methods are combined, results reflect the
additive benefit of each.
Many of the strengths of the results of this chapter carry over from the challenging nature
of the stimuli set created. To reiterate Chapter 1, the image pairs in the face-matching
task consisted of true match (two different images of the same celebrity) and false match
(one image of a celebrity and one of a lookalike for that celebrity). The stimuli used in this
chapter are pixelated, making a task that addressed a difficult version of a practical
problem. This design allowed us to test the improvements to face recognition in a more
challenging situation than ever before. In the past, ceiling effects had haltered
improvements. The more difficult task that I created allowed the effects of combinations
of techniques to be recognised, as base level performance was much poorer than on
previous tasks, the improvements made by techniques used alone and in combination
could be measured as even combined techniques did not bring performance to ceiling on
the pixelated celebrity lookalike face matching task.

I demonstrated that combining techniques helped to improve results, even more than
when any one technique was used in isolation. However due to the limited testing
opportunity with the super-recognisers, I do not have results for the super-recognisers on
the blurred pixelation task, and consequently could not perform crowd analysis on
blurred pixelation data for super-recognisers. To more fully understand the effect of
combining the three improvement methods I investigated, it would be interesting to test
whether performance could be improved even more through all three techniques applied
in all combinations i.e. testing the SRs on the blurred version of the task and then in
addition to this applying crowd analysis to the results.

Each of these techniques, and the combinations addressed in this chapter could be used
in applied scenarios to achieve more accurate face identity decisions. Blurring, crowd
analysis and the performance of super-recognisers provide ways of improving
performance where past methods of training have failed. These techniques are far easier

118

and less time consuming to implicate than training, furthermore crowd analysis has the
unique advantage of being applicable to improve performance levels using pre-existing
data. These methods improve face-matching performance without relying on familiarity
with the faces that are available for comparison, making them particularly useful in
applied situations where familiarity with the faces concerned is not always a viable
option.

Up until now, this thesis has focused on face recognition for challenging stimuli based on
incidental differences in appearance across multiple images of a face and incidental
similarity in appearance between different person trial images. This has shown that whilst
it is possible to incidentally look like somebody else, it is also possible for images of the
same person to incidentally look like different identities across images (e.g. Jenkins et al.
2011). All images investigated so far have been ambient in nature, however there could
be particular images changes which cause these situations to occur. In particular, the
distance from which an image is taken from may influence the appearance of a face, and
in turn influence the perceived identity. Chapter 4 will explore this effect of changing
camera-to-subject distance, looking at both the effect on the configural information in a
face as portrayed in an image, and whether any changes of configuration effect face
matching ability for unfamiliar and for familiar viewers.

119

Chapter 4 – Changing Camera-to-Subject Distance
4.1 Chapter Summary
The experiments in this chapter investigated effects of camera-to-subject distance on
configural properties of the face image. Changes in camera-to-subject distance produce
non-linear changes in face measurements across images. These changes reduced accuracy
in an unfamiliar face-matching task by making same identity images look less like
themselves. Identity matching performance was much poorer when unfamiliar viewers
compared photographs taken from differing distances, than when the comparison images
were taken at the same distance. Familiar viewers were far less affected by this distance
change and performed at very high accuracy levels in both conditions. Distance cues
compensate camera-to-subject change, suggesting the operation of perceptual constancy
mechanisms in the high level visual domain of face shape.

4.2 Introduction
The earlier chapters of this thesis, in line with previous research, have shown that facematching performance is generally poor for unfamiliar viewers. Challenging images in the
foregoing experiments (Chapters 2 & 3) resulted in even poorer performance than
previous experiments based on cooperative stimuli. So far in this thesis I have focussed
on challenging image performance for ambient face images. These images help to capture
natural variability in the appearance of any given face. For example, facial expression and
pose may change across images, and the environment that face photographs are taken in,
can differ. The effects of such factors have previously been examined in isolation, and it is
well reported that these superficial image changes can result in impaired performance on
tasks involving identity judgment (Bruce, 1982, 1994; Johnston, Hill, & Carman, 1992;
Troje & Bulthoff, 1996; O’Toole, Edelman & Bulthoff, 1998; Bruce et al. 1999) However,
one interesting and potentially important factor that has received little attention is the
effect of camera-to-subject distance on identity judgments.

120

A few isolated findings suggest that changes in camera-to-subject distance may well
affect identification accuracy. In an interesting paper on optics, Harper & Latto (2001)
photographed five models (two male, three female) from five different distances (0.32m,
0.45m, 0.71m, 1.32m & 2.70m) and standardised the size of the resulting images (see
Figure 4.1). Face images taken further away were visibly flatter, giving the models a
heavier appearance and the implication is that the faces appeared to have different
shapes in these different distance conditions. Participants gave higher estimates of the
models’ weight as camera-to-subject distance increased. In a later study (Bryan, Perona &
Adolphs, 2012) ratings of trustworthiness, competence, and attractiveness were lower
when the camera-to-subject distance was reduced (specifically when the photographs are
taken within personal space). Taken together, these findings suggest that facial
appearance changes as a result of camera-to-subject distance, yet no studies have looked
at how these changes affect performance in tasks involving identity matching judgments.
This omission is perhaps surprising, given the emphasis on configuration in the face
recognition literature. The configural processing account holds that each face has a
unique configuration that is learnt, and that it is knowledge of this configuration that
allows us to tell familiar faces apart. This position seems to require that the configuration
of a particular face stays constant across images, but the findings of Harper & Latto (2001)
and Bryan et al. (2012) suggest that it does not. Given that camera-to-subject distance is
rarely kept constant in practical applications of identity matching, effects of camera-tosubject distance on face identification would also be interesting from an applied
perspective.

Figure 4.1 Changes in face shape resulting in differing weight judgments as photographs were taken from
far distance (left) to near distance (right), example taken from Harper & Latto (2001).

121

In a compelling illustration of this Burton et al. (2015) assessed the stability of facial
configurations by measuring distances between features in multiple images of three
famous politicians (David Cameron, Barack Obama and Angela Merkel), captured from
unknown, but presumably different distances (see Figure 4.2 for example). They found
that distances between features varied as much between photos of the same person as
between photos of different people. This observation seems to challenge a
straightforward configural processing account of face recognition and face learning.
However as the camera-to-subject distance for these examples was unknown, a more
formal assessment of its effects was not possible. Moreover, as camera-to-subject
distance was not the exclusive focus of that study, variation in other factors such as pose
could also have affected the configural measurements.

Figure 4.2 Example of measurement figure taken from Burton et al. (2015). Images are standardised so that
interocular distance is the same. Metric distances are expressed as proportions of standardised interocular
distance.

122

It is important to note that the politicians were easily identifiable to familiar viewers in
any of the face images, despite the fact that these images varied in configuration. It is a
well established finding that unfamiliar face recognition can be derailed by superficial
image changes, whereas familiar face recognition is robust against such changes (Burton
et al. 1999; Jenkins et al. 2011). This contrast has been linked to familiar viewers having
more perceptual experience with the range of appearances that the face can take
(Jenkins et al. 2011). Assuming that familiarity interacts with distance-related changes in a
similar fashion, it seems likely that these too will impair performance more for unfamiliar
viewers than familiar viewers.

Although no previous studies have examined effects of camera-to-subject distance in an
identity matching task, a study by Liu (2003) investigated the effect of camera-to-subject
distance on face recognition accuracy in a memory task. A face image of each identity
used as stimuli in this experiment was edited using Matlab software to create two
versions of each face image – one which reflected the appearance that the face would
take at a far distance, and one that reflected the appearance of the face from a near
distance. Participants viewed face images in a learning phase, and at test were shown
either the identical image, or the image altered to reflect the distance which differed
from the test image. Liu (2003) showed that the same image of a face is harder to
recognise as having been seen before if the image is digitally altered to reflect the
changes that would result from altering camera-to-subject distance. My question of
interest is different to this in at least two important ways. First, manipulating a single
image of a face is not the same as presenting two different images of a face (Bruce, 1982;
Jenkins et al. 2011). There is both behavioural (Bruce, 1982) and neural evidence for this
(Bindemann, Burton, Leuthold and Schweinberger, 2008). Bindemann et al. (2008)
reported that the brain responses from the N250r (an event related brain response
generated from the fusiform gyrus) differ depending on whether multiple images of the
same face are presented, or if repetitions of the same face image are viewed. Perhaps
more importantly with reference to Liu’s (2003) study, the N250r response was the same
for repetitions of the same face image and for when a digitally altered version of this

123

image (stretched) was presented. In addition to this, fMRI adaptation has shown that
neural representations in the Fusiform Face Area (FFA) to images of the same face does
not vary with changes in size, expression or pose of the face in the image (Grill-Spector et
al. 1999; Andrews & Ewbank, 2004). These results highlight that responses to stimuli may
differ when simply altering the same image of a face digitally, compared to presenting
multiple images. Liu’s stimuli really created an image recognition task, as images were
changed only by simulated distance manipulation.

Second, I am interested in perceptual matching, as distinct from recognition memory.
Perceptual matching is interesting in its own right because it allows us to set aside aspects
of task difficulty that arise from the fallibility of memory, and to focus on those difficulties
that remain at the perceptual level. Face matching tasks also model the task faced by
forensic and security officials of determining identity from multiple face images.

I will conduct 3 studies to investigate the effects of changing camera-to-subject distance
on facial appearance in an image, and the consequent effects of those image changes on
tasks involving identity judgment. I will first characterise configural changes across
multiple images of the same individuals taken at known distances. I will then investigate
whether these changes translate into difficulties in face-matching accuracy for familiar
and unfamiliar viewers. Finally I will evaluate whether the visual system compensates for
distance related changes in the face image when distance cues are available.

4.3 Experiment 7: Facial Configuration Measurements
The purpose of this study was to relate camera-to-subject distance to facial configuration.
The apparent size of an object changes with viewing distance, in the sense that the size of
the retinal image changes. Linear changes in the size of a face image (e.g. rescaling a

124

photograph) do not affect configural information because they do not alter the relative
distances between features. Consistent with this observation, both behavioural and
neuroimaging studies have found that face recognition is unaffected by linear rescaling
(Grill-Spector et al. 1999; Andrews & Ewbank, 2004). For 3D objects (e.g. live faces as
opposed to face photographs), the optical situation is different. Changes in camera-tosubject distance generate non-linear changes in the image, such that different parts of
the image are affected to differing degrees (Pirenne, 1970). For convex objects, including
faces, distant viewing leads to flatter appearance, whereas closer viewing leads to more
convex appearance (see Figure 4.1). To relate this transformation to the notion of
configuration in the face perception literature, I measured distances between key facial
features in photos that were taken at different viewing distances. My expectation was
that, as a reflection of the flat-to-convex variation, the change in viewing distance would
affect measures near the edge of the face more strongly than it affects measures near the
centre of the face.

Photographic Procedure

The images used for all of these studies were face photographs of 18 final year
undergraduates at the University of York. To allow construction of face-matching
experiments (Megreya & Burton, 2008; Burton et al. 2010) these models were
photographed in 2 separate sessions, one week apart. In each session, each model was
photographed at 2 distances—Near (camera-to-subject distance = .32m) and Far (camerato-subject distance = 2.7m), following Harper & Latto (2001). This resulted in 4
photographs for each of the 18 models: Week 1 Near, Week 1 Far, Week 2 Near, and
Week 2 Far (72 photos in total). All models were photographed with a neutral expression
using an Apple iPhone 5 on default settings. Photos were then cropped around the head
to remove clothing and background. For anthropometric analysis, all images were resized
to an interocular distance of 150 pixels, preserving aspect ratio (see Figure 4.3).

125

Anthropometric Analysis
For each model in each condition, I measured 5 feature-to-feature distances that have
been specified in the configural processing literature: corner of eye to edge of nose (left
and right; Leder & Carbon, 2006), corner of nose to corner of mouth (left and right; Leder
& Bruce, 2000), and nose to mouth (Leder & Carbon, 2006). Precise anatomical definitions
were as used by Burton et al. (2015). The corner of the eye is defined as the centre of the
canthus, the corner of the nose as the lateral extent of the nasal flange, and the corner of
the mouth as the lateral extent of the vermillion zone. Figure 4.3 shows these
measurements for one model. All distances are expressed in units of standardized
interocular distance.

Figure 4.3 Example of the measurements taken for two of the photos of one model. Measurements taken
were the distances between: left eye to nose, right eye to nose, left nose to mouth corner, right nose to
mouth corner and centre of nose to mouth.

Results and Discussion
For each of the 5 feature-to-feature metrics, I performed a separate 2x2 ANOVA with the
within-subjects factors of Photographic Session (Week) (Week 1 versus Week 2) and
Camera-to-Subject Distance (Distance) (Near versus Far). Results of these analyses are
summarized in Table 4.1.

126

Week

Dist

Avg1

Avg2

Diff

F

p

ES

AvgN

AvgF

Diff

F

p

ES

EN(L)

.57

.57

.00

.05

>.05 .00

.58

.56

-.02

1.67

>.05

.09

EN(R)

.58

.57

.01

.20

>.05 .01

.59

.57

-.02

3.53

>.05

.17

NM(L)

.42

.41

-.01

1.58 >.05 .09

.38

.45

.07

39.82 <.001 .70

NM(R)

.42

.41

-.01

1.23 >.05 .07

.38

.45

.07

42.97 <.001 .72

NM(C)

.22

.23

.01

1.10 >.05 .06

.20

.24

.04

17.26 <.001 .5

Table 4.1 Table showing mean measurements for each photograph condition. EN stands for ear to nose
measurement, and NM represents nose to mouth. The letters following denote the side of the image which
the measurement was taken for, L = left, R = right & C = centre. Average measurements are calculated for
week 1 (Avg1) and week 2 (Avg2) at both near (AvgN) and far (AvgF) distances.

Photographic Session (Week) had no significant effect on any of the measurements (p > .1
for all), indicating similar viewpoint and expression in both sessions. More importantly for
this study, Camera-to-Subject Distance (Distance) had a significant effect on some
measures but not on others. Specifically, the more peripheral nose-to-mouth
measurements were greater for Far images than for Near images, whereas the more
central eye-to-nose measurements were statistically equivalent at the two camera
distances we compared. This pattern in the anthropometric data corroborates the flatter
appearance of the Far images and the more convex appearance of the Near images. More
generally, it confirms the non-linear effect of camera-to-subject distance on configural
information for this image set: some measurements change more than others. I next used
a paired matching task to assess the implications of these configural changes for
perception of facial identity.

4.4 Experiment 8: Face-Matching & Camera-to-Subject Distance
In the GFMT, a standard matching experiment, participants are presented with pairs of
face photographs that were taken with different cameras (Burton et al. 2010). For each
pair, the participant’s task is to decide whether the 2 photos show the same person
(Same trials; 50% prevalence) or 2 different people (Different trials; 50% prevalence).
Despite the simplicity of this task, error rates are high when the faces are unfamiliar, as

127

the viewer has no way to distinguish image changes from identity changes. When the
faces are familiar errors are virtually absent, presumably because variation in appearance
is better characterised by the viewer (Jenkins & Burton, 2011).

Here I extend the standard paired matching task by adding camera-to-subject distance as
an experimental factor. Because face images change with camera-to-subject distance,
manipulating this distance allows very specific predictions to be made: for viewers who
are unfamiliar with the faces concerned, a change in camera-to-subject distance should
impair performance on Same Identity trials (because it generates dissimilar images) and
should improve performance on Different Identity trials (for the same reason). If identity
judgments by familiar viewers rely on facial configurations then the same should apply to
their performance. However, given that familiar viewers readily see through changes in
viewpoint, lighting, facial expression, and other factors, it is anticipated that familiar
viewers might similarly see through changes in camera-to-subject distance, such that
their performance would be unaffected by this manipulation

Method

Participants
45 psychology undergraduates at the University of York participated in exchange for
payment or course credit. 23 of these participants were first-year students who arrived at
the University of York after our photographic models had left, and hence had never seen
the faces in the stimulus set (verified post-test; see Procedure section below). We refer to
these participants as Unfamiliar viewers (M = 4, mean age = 18.7). The remaining 22
participants were other students from the same year group as our photographic models,
and had spent over two years studying on the same course (M = 3, mean age = 22.14). I
refer to these participants as Familiar viewers because they had seen the faces in the
stimulus set routinely over those two years.

128

Stimuli and Design

The same stimulus images were used as in Experiment 8. Images were cropped around
the face to remove extraneous background. The face images measured 700 pixels wide by
900 pixels high.

In order to create the face-matching task, images were paired according to Same and
Different Identity trials. The different identity trials were created by pairing the most
similar face images from those available. Image pairs were also constructed according to
Same and Different Camera-to-Subject Distance. Crossing the within subject factors
Identity and Distance resulted in four stimulus conditions: (i) Same Identity, Same
Distance; (ii) Same Identity, Different Distance; (iii) Different Identity, Same Distance; and
(iv) Different Identity, Different Distance. Figure 4.4 shows example pairs from each
condition.

Figure 4.4 Example of one identity with each of their four image identity pairings shown. The first column
shows image pairs of the same identity and the second column shows different identity pairs. The first row
shows same camera-to-subject distance pairs and the bottom row shows pairs where the images are of
different camera-to-subject distance.

129

For each pair, the participant’s task was to decide whether the two images showed the
same person or two different people. This allowed a percentage accuracy score to be
calculated for each participant in each condition. Both Unfamiliar and Familiar Viewer
groups carried out this task.

Procedure

Participants were tested in groups and worked individually in silence. Participants viewed
a total of 144 image pairs as part of a face-matching task. These images were viewed as
part of a timed PowerPoint presentation in which each image pair was visible for 5
seconds followed by a 3,2,1 countdown before the next image pair appeared. Within this
viewing and countdown time participants were required to decide whether the images in
the pair they had just viewed showed the same person’s face or if the faces were two
different people. Participants recorded their answers by circling ‘same’ or ‘different’ on
an answer sheet to indicate their identity judgement for each image pair.

After completing the face-matching task, participants completed a familiarity check. For
this, participants viewed images of each of the identities that were stimuli in the task and
simply ticked a box to indicate whether they were familiar with each face or left the box
unmarked if the faces were unknown to them. This allowed me to exclude the data from
faces that were unfamiliar to any of the familiar viewer group, or familiar to any of the
unfamiliar group.

Results

A 2x2x2 mixed ANOVA between Identity, Viewer Familiarity and Distance was performed
and confirmed that overall performance was significantly lower for Unfamiliar viewers (M
= 85.71, SE = .80, CI = 84.11 – 87.31) than Familiar viewers (M = 97.89, SE = .81, CI = 96.25
– 99.53), F(1,43) = 114.8, p<.001, η p2 = .73, this finding was expected given that familiar
viewers have been found to be more robust against superficial image changes than
unfamiliar viewers. There was a significant main effect of Identity, with performance

130

being higher for Different Identity (M = 93.36, SE = .91, CI = 91.53 – 95.2) than Same
Identity trials (M= 90.25, SE =.87, CI = 88.49 – 92.00), F(1,43) = 5.17, p = .03, η p2 = .11.
However there was no interaction between Identity and Viewer Familiarity [F(1,43) =
1.53, p =.22, η p2 = .03].

There was also a significant main effect of Distance, with participants performing better
for Same Distance trials (mean = 94.92, SE =.55, CI = 93.82 – 96.03) than Different
Distance trials (mean = 88.69, SE = .85, CI = 86.97 – 90.38), F(1,43) = 51.90, p<.001, η p2
=.547. This demonstrates that changing camera-to-subject distance lowers face-matching
performance accuracy. In addition to this both the interaction between Distance and
Viewer Familiarity [F(1,43) = 21.80, p<.001, η p2 = .34] and the interaction between
Distance and Identity [F(1,43) = 133.64, p<.001, η p2 =.76] were significant. Finally, there
was a significant three way interaction between Identity, Distance and Viewer Familiarity
F(1,43) = 117.45, p<.001, η p2 = .73. To break down this 3-way interaction, I next carried
out separate 2x2 within-subjects ANOVAs for the Familiar and Unfamiliar groups.

131

Figure 4.5 Effect of changing camera-to-subject distance on performance accuracy in the face-matching task
for familiar (F) and unfamiliar viewers (U), for same and different identity trials, at both same and different
distances. Error bars show the standard error of the mean.

As expected, Familiar viewers performed with very high accuracy in both the Same
Distance and Different Distance conditions. Accuracy was significantly higher for Same
Distance image pairs (M = 98.84 % correct, SD = 1.37) than for Different Distance image
pairs (M = 96.70 % correct, SD = 4.17), [F(1,21) = 5.49, p<.05, η p2 = .21], even though the
magnitude of this effect was small. There was no significant effect of Identity [F(1,21) =
3.41, p = .08, η p2 = .14] and no interaction between these two factors p>.05.

More importantly, Unfamiliar viewers performed significantly better when the paired
images were taken at the Same Distance (M = 90.85% correct, SD = 4.97) compared to
Different Distance (M = 80.54, SD = 6.83) [F(2,22) = 51.20, p<.001, η p2 = .67]. There was no
significant effect of Identity [F(1,22) = 3.48, p=.08, η p2 = .14], but critically there was a

132

significant Identity Distance interaction [F(1,22) = 155.13, p<.001, η p2 = .88]. For Same
Identity pairs, accuracy was higher for Same Distance (M = 97.33, SD = 3.09) images than
Different Distance (M = 69.29, SD = 13.41) images. For Different Identity pairs the
opposite was true: accuracy was higher for Different Distance (M = 91.86, SD = 6.29)
images than for Same Distance images (M = 84.37, SD = 10.64). Simple main effects
revealed that there was a significant effect of Identity for both Same [F(1,44) = 19.34,
p<.001, η p2 =. 31] and Different Distance images [F(1,44) = 58.73, p<.001, η p2 =. 57]. And
there was a significant effect of distance for Same Identity pairs [F(1,44) = 191.92,
p<.001, η p2 =. 81], and also for Different Identity pairs [F(1,44) = 13.69, p<.001, η p2 =. 31].

Discussion

Changing camera-to-subject distance impaired viewers’ face-matching ability. Accurate
face recognition remained easy for viewers who were familiar with the faces concerned
regardless of camera-to-subject distance (mean performance across same distance and
different distance conditions = 99% accuracy). Unfamiliar viewers performed much more
poorly in the different distance condition (M = 81%) than the same distance condition (M
= 91%). There was a crucial cross over interaction in the unfamiliar group – with
performance decreasing as a result of changes to camera to subject distance only for
same identity trials.

This pattern of results for unfamiliar viewers suggests that the kinds of configural change
evident in the measurement study join the long list of image changes that can thwart
identification. This finding shows that the effects of changing camera-to-subject distance
identified in the past (Latto & Harper, 2007) do indeed also impair face recognition.
Distance related image changes had been found to impair face memory for digitally
edited compared with identical images (Liu, 2003). My study demonstrates that distance
manipulations also impair face-matching.

133

The finding that familiar viewers were impervious to these non-linear changes in facial
configuration suggests that familiar face recognition is not strongly dependent on
distances between features in face recognition. It seems that when learning a face,
people are not learning any specific configurations of a face, as familiar viewers are able
to see through changes in camera-to-subject distance, where these configurations differ
across the images.

The findings of these studies raise an interesting question: if faces look more flat or
convex at different viewing distances, why do we not notice these shape changes in daily
life? The next experiment addresses this question.

4.5 Experiment 9 - Perceptual Constancy for Face Shape
In the measurement study I demonstrated that images undergo non-linear changes in
configuration as a result of changed distance. Experiment 8 showed that these image
changes can easily disrupt unfamiliar face matching: unfamiliar viewers were poorer at
matching pairs of faces when the two images were taken from different camera-tosubject distances, compared with when the images were taken from the same distance.
In real life however, we do not tend to notice changes in face shape. For example, faces of
people walking towards us do not appear more convex as they approach. There are in fact
lots of examples of not noticing image change for images other than faces. These include
changes in colour and brightness, for example clothes are perceived as the same colour
even under different types of lighting; the colour is perceived as constant through
calibration of white (Webster & Mollen, 1995). Shapes are also perceived constant
through the use of depth cues (Pizlo, 1994). Each of these scenarios is an example of
perceptual constancy – where the visual system uses information from the environment
to overcome image changes. Perceptual constancy has been studied intensively but
normally in the context of low-level visual features such as colour and shape.

134

It is evident that perceptual constancy can help viewers make sense of the colours or
shapes that they see. Following this, one possibility is that perceptual constancy
mechanisms compensate for changes in face shape caused by viewing distance. In the
same way that viewers perceive the shape of an open door as rectangular, even though
the retinal projection is trapezoidal, they may view face shapes as a constant shape
despite distance-related distortions by taking account of the viewing distance for the face
image. It may be that when looking at a photograph on a screen no information of
viewing distance is available, removing the ability to compensate for distance and as a
result the face images look different.

In order to investigate if perceptual constancy applies to face shape under these
conditions I will investigate whether a distance cue, indicating that one image was taken
from further away and one close up, can overcome the effect that a change of camera-tosubject had on recognition accuracy as found in Experiment 8. My approach to this will be
to manipulate congruency between the actual camera-to-subject distance and the
distance implied by cues in the display. Specifically, I will present the two face images in
each pair at two different sizes – a small image implying a long viewing distance and a
large image implying a short viewing distance. In congruent trials, the size cues will be in
sympathy with the images, so that the near image is larger and the far image is small. In
incongruent trials, the size cues will conflict with the images, so that the near image is
small and the far image is large. If a perceptual constancy mechanism compensates for
distance-related distortions, then participants should perform more accurately on
congruent trials than on incongruent trials, when the two images show the same person.
This is because the valid distance cues will allow constancy to correctly ‘undo’ the optical
distortion, making the images look more similar. At the same time, accuracy should be
lower for congruent trials than for incongruent trials, when the two images show
different people. This is because with invalid distance cues, the constancy mechanism will
compensate in the wrong direction, exaggerating differences between the images,
making them look less similar.

135

Method
Participants
30 unfamiliar participants (male=6, age = 19) from the University of York took part as
participants in this study.
Design
Experiment 9 was largely based on the design of Experiment 8. However, in this
experiment the image pairs featured a distance cue. In half of the trials this distance cue
was Congruent with the true distances for which the photos were taken e.g. the near
photo was shown as a bigger image on screen and the far image as smaller. Other times I
swapped the image sizes, to create an Incongruent image pairs e.g. the near image was
made smaller to appear far away, and the far image enlarged to appear near.

Congruent image pairs were created by keeping the near photograph its natural size (the
original size of the saved photograph, not resized in any way), and the far image its
natural size. This resulted in a larger ‘near’ photograph, and a smaller ‘far’ photograph.
Incongruent images were created by resizing the far image to be the natural size that the
near image was, and resizing the near image to take the size that the original far image
was. This created a small ‘near’ image and a large ‘far’ image, hence incongruent to the
natural display format these images would take. In addition to this, perspective lines were
added to the images to provide an additional depth cue. The perspective lines supported
the interpretation of distance in the displays. See Figure 4.6 for an example of congruent
and incongruent same and different identity pairs.

136

Figure 4.6 Example of congruent and incongruent face image pairs (with distance cues) for same and
different identities.

Procedure
The task was a face-matching task as in Experiment 8, however the stimuli now featured
distance cues. Participants were tasked with making same or different identity judgments
for each face pair that they saw. 144 pairs were viewed in total.

137

Results & Discussion

Figure 4.7 Graph showing the percentage of correct responses for both congruent and incongruent image
pairs broken down by same and different identity trials. Error bars show the standard error of the mean.

A 2x2 ANOVA was conducted to compare the effect of Distance Cues (Congruent &
Incongruent) and Identity (Same & Different), on face matching accuracy (see Figure 4.7).
The analysis revealed no significant main effect of Distance Cue, comparing accuracy on
Congruent (M = 80.82) with Incongruent (M = 78.49) trials F(1,27) = 3.29, p = .08, η p2 =
.11. However, there was a significant main effect of Identity F(1,27) = 8.13, p = .008, η p2 =
.23, with higher recognition accuracy for Different (M = 85.35) rather than Same (M =
73.97) identity trials. More importantly, there was also a significant interaction between
Congruence and Distance, F(1,27) = 23.22, p<.001, η p2 =.46, confirming that the
congruency manipulation had opposite effects for Same Identity and Different Identity
trials.

138

Simple main effects revealed that as anticipated, participants were significantly more
accurate at Same identity trials for Congruent image pairs (M = 78.38, SE = 3.3) than
Incongruent image pairs (M = 69.57, SE = 3.6), F(1,54)= 22.44, p<.001, η p2 = .29.

My second prediction was also met with results showing that for different identity trials,
viewers were more accurate for incongruent image pairs (M = 87.42, SE = 1.8), than for
congruent image pairs (M = 83.27, SE = 2.18), F(1,54) =4.97, p<.01, η p2 = .08.

The results from Experiment 9 suggest the operation of a perceptual constancy
mechanism at the level of face shape. For unfamiliar faces, participants were better for
same identity trials when provided with a congruent distance cue. In this situation the
constancy mechanism would be giving the information necessary to make the images
look more similar e.g. compensating in the correct direction for the differences in images
as a result of distance.

Further evidence for the constancy mechanism comes from the result found for different
identity trials. In the case of incongruent different identity face trials if the logic applied
above was followed, the incongruent distance cues would lead to being compensated in
the ‘wrong’ direction, making the images look even more different than they would
otherwise. Hence correct different judgments were greater when the differences were
amplified due to incongruence than when they were viewed in the congruent condition. If
no perceptual constancy mechanism were being applied to face perception no differences
in performance would have been observed across conditions.

139

4.6 General Discussion
In this chapter I have shown that changes in camera-to-subject distance lead to non-linear
changes in face measurements across images. These changes had a detrimental effect on
face identification efforts of unfamiliar viewers – identity matching performance was
much poorer when unfamiliar viewers compared photographs taken from differing
distances, than when the comparison images were taken at the same distance. This effect
was driven by difficulty for matching same identity trials. Familiar viewers were far less
affected by this distance change and performed at very high accuracy levels in both
conditions. Furthermore I provide evidence of perceptual constancy effects at the level of
face shape. Congruent distance cues aided recognition of same identity faces, with
distance cues compensating for the apparent differences in faces due to camera-tosubject distance. In line with this, an incongruent image cue made the images look even
more different to each other, and hence increased accuracy on correct different person
judgements. These findings suggest that perceptual constancy can account for the
apparent continuity of facial appearance at different viewing distances, and that valid
distance cues are required for the smooth operation of this mechanism.

Previous research had shown that perceptions gathered from a face image, including
weight estimates and social inferences, differed as a result of changed camera-to-subject
distance (Harper & Latto, 2001; Bryan et al. 2012). My analysis showed that non-linear
changes in metric distances between facial features can account for such perceptions.
Most importantly, in addition to the social inference effects reported previously, my study
demonstrated that these configural changes affect accuracy in identity judgment. The
camera-to-subject manipulation greatly impaired unfamiliar viewers, whereas familiar
viewers were barely familiar at all. The level of familiarity with each face was not
recorded – participants rated faces as either familiar or unfamiliar. It could be the case
that some of the familiar viewers were not all that familiar with the faces involved, which
could explain the small difference found for familiar viewers over distance change. My
findings are very much in line with past research demonstrating superior performance of
familiar viewers in face recognition task over unfamiliar viewers (e.g. Burton et al. 1999,

140

Jenkins et al. 2011). In addition this research is the first to show that face matching
performance, in addition to face memory (Liu, 2003) is impaired by camera-to-subject
distance change across images.

Moreover, the findings of this chapter suggest that learning configural information – at
least in the conventional sense of distances between features – is not the key to learning
a face. Experiment 7 highlighted that the configuration of a face does not remain constant
across multiple images if these images have been taken from different camera-to-subject
distances. Yet, it is known that people can accurately identify celebrities who have
become familiar faces through our exposure to images and video footage, which capture
the celebrity from many different distances. It seems more likely that when people are
become familiar with a face they are gaining experience under a variety of conditions, this
would include learning how the face looks from a range of different distances (e.g.
Jenkins, 2011). This fits with the conception of face space advanced throughout this
thesis. On this account, familiar viewers have a range of experience with a face, and
hence experience of seeing the face over several different distance configurations.
Unfamiliar viewers do not have this refined face space and must instead make identity
judgments using only the information that is in the images in front of them. This lack of
perceptual constancy interacts with configural change in the following ways. First it
makes it harder to ‘tell together’ same identity images with different configurations, as
the viewer does not have experience of the variability in appearance for that identity.
Second, it supports viewers’ performance on different identity trials, as the different
camera-to-subject distances tend to exaggerate natural differences between the faces
such that the identities would be more likely to be categorised as belonging to different
identities in face space. Notably, the distance change only makes different people look
more different because they look similar to begin with - it would be possible for different
people to look less different as a result of distance change if for example one was fat and
one was thin.

141

These experiments also advance the theoretical understanding of unfamiliar face
recognition; in particular I have proposed that a perceptual constancy mechanism exists
whereby distance cues help an individual make sense of a face image viewed. Whereas
past studies have shown perceptual constancy for colour and basic geometric shapes, and
even for the size and shape of humans, to my knowledge, this is the first study to propose
a constancy mechanism for faces.

Several practical implications follow from these findings. Firstly, according to the results
of Experiment 7, anthropometry (in which the images used in investigation are likely
taken from unknown different distances) is not a reliable method of facial identification.
This is because images of faces do not hold constant configurations - measurements
between features of a face change across images when these images have been taken
from different distances (Kleinberg, Vanezis & Burton, 2007). Additionally, as the results
of Experiment 8 show that changing the distance from which a photograph of a subject
was taken between images reduces identification accuracy when making identity
judgments based on face-matching, it would be advised that wherever possible, there
should be consistency in distance from which photographs are taken from in forensic or
security situations. For example better identification rates may be met if security officials
took photographs of a suspect from a standard distance, which could be compared with
photographs taken of the same suspect, from this same standard distance, following a
separate incident, no matter where the incident occurred. If this is not possible, caution is
urged when comparing images that were taken at different or unknown distances,
because there are systematic differences in how they will appear. Familiar viewers are
largely exempt from this caution because they can see through configural changes easily.
Where this is not possible, providing accurate distance cues could improve performance
for unfamiliar viewers.

In conclusion, face configuration in an image changes as a result of changes in camera-tosubject distance. These changes affect performance in face-matching tasks, with
unfamiliar viewers being very strongly affected. The finding that accurate distance cues

142

help viewers to see through these changes in configuration suggests the operation of a
perceptual constancy mechanism.

Up until now this thesis has addressed incidental creations of, or naturally occurring,
challenging stimuli. The next chapters of this thesis will move on to explore a second type
of challenging stimuli, where deliberate attempts are made to change appearance
through disguise.

143

Chapter 5 – Matching Disguised Faces
5.1 Chapter Summary
I created a sophisticated disguise face database which is the first to include both evasion
and impersonation disguise items. Models were unguided in their disguise efforts. Any
props that they requested to aid their disguises were provided and all models were
informed of the effect of changing face shape as a result of camera to subject distance
(Chapter 4).

Disguise impaired face-matching performance. Performance for unfamiliar viewers was
poor, and knowing that disguises may be present did not improve this. Evasion disguise
caused most difficulties for face-matching, followed by ‘impersonation similar’ then
‘impersonation random’ disguises. Familiar viewers were much better at seeing through
disguise than unfamiliar viewers but even familiarity did not help viewers completely see
through evasion disguise. Links to theories of face space and familiarity are discussed.

5.2 Introduction
This chapter will explore face-matching performance for disguised faces. Previous
research has shown that unfamiliar face-matching is poor, with people making between
10-20% errors on the GFMT (long & short version respectively) which is a standardised
test of face image matching, and also in matching tasks involving an ID card image and a
physically present person (Burton et al. 2010; Kemp et al., 1997; White et al., 2014).
However these results come from tasks that have used cooperative stimuli (when same
identity trials were created there was no deliberate attempt to make an individual look
different across multiple images) and mismatch trials were created from the most similar
match taken from a limited sample of face images.

144

I found that performance was even worse when using more difficult stimuli images
involving incidental disguise (Chapters 2 & 4) than had been found previously for
cooperative stimuli. These include the case of ambient same person images and
extremely similar imposter face images (Chapter 2) and images where face shape is
changed as a result of camera-to-subject distance (Chapter 4). These studies have helped
in gaining a better understanding of face recognition in challenging circumstances and my
findings have also demonstrated the power of familiarity – so far familiarity has led to
large increases in performance on the difficult tasks that I have developed.

With such promising results, I now invest in creating a disguise face database to allow
face-matching performance accuracy to be tested in what may be the most challenging
case yet – when a person is deliberately trying to evade their own identity or impersonate
someone else. As I am creating the database from scratch it will include all aspects of
disguise that I am interested in. Past studies have relied on incidentally occurring
similarities between faces to provide different person trials in face-matching tasks and
used two images of the same person that may differ in naturalistic incidental ways. Now,
in addition to creating no disguise image pairs similar to those used in past matching
tasks, I attempt to create disguise image pairs. These will include creating deliberate
similarities between different identities (impersonation disguise) as well as creating same
person trials where there is a deliberate attempt to alter appearance across images
(evasion disguise). These images will allow face-matching performance for disguised faces
to be directly compared to performance for undisguised faces for the first time.

Disguise in Face Recognition Memory
Only a few past studies have attempted to approach questions related to face recognition
performance and disguise. Patterson and Baddeley (1977) were the first to do so,
publishing a paper that consisted of two disguise face memory experiments. Their first
experiment tested face recognition memory performance for identical face images
compared with face images changed in pose or expression, and also faces in disguise. The

145

disguise stimuli were images of actors, with the same actors photographed in two
different character roles. This allowed one of these images to be presented during the
learning phase and the other at test. Participants were divided into two groups for the
learning phase in the experiment. Both groups were required to view face images with
the intent of memorising them. However, one group was instructed to focus on making
personality judgments for the faces they saw and the other instructed to focus on the
features of the faces (discussed further in Chapter 5). Participants were aware that the
appearance of the individuals might change when they saw the face again at test. At test
participants again viewed face images, and were this time tasked with deciding whether
or not the identities presented at test had featured in the earlier memory task.
Recognition performance was poorer for images presented in disguise at test than images
that were presented unchanged [identical image presented again at test] or changed only
in pose or expression. Recognition performance was near chance for faces presented in
disguise at test.

The disguise stimuli used in Patterson & Baddeley’s (1977) first experiment in their paper
discussed above, were multiple images of the same actors, whose appearances differed
across images as a result of matching different character roles for different acting jobs.
This stimuli acquisition method meant that the exact changes made to appearance by the
actors across images were unknown to the experimenters, and hence the effect of each
change could not be easily investigated. A more controlled manipulation of disguise was
applied in Experiment 2. Here Patterson & Baddeley created the disguise stimuli
themselves by asking volunteer male models to create disguises using props. Disguise
through purely the addition of props is a common approach to disguise undertaken by
past work. I will here on refer to disguise by props alone as simple disguise. Models in
Patterson & Baddeley’s study were instructed to add wigs, add or remove beards, and
add or remove glasses across a series of photographs. I will refer to disguise based on the
experimenter’s instruction as prescribed disguise and situations where models disguise
themselves as they wish as free disguise. Patterson & Baddeley’s experiment thus fits the
category of prescribed simple disguise. Participants were tested on their recognition
performance for disguise compared to no disguise images in a similar manner to in

146

Patterson & Baddeley’s Experiment 1. As in the case of the recognising disguise actor
images, participants in Patterson & Baddeley’s (1977) second experiment were worse at
recognising a face identity when it was presented in a different disguise at test to
learning, than recognising an exact image that they had viewed before. Specific effects of
each disguise manipulation investigated in this experiment are discussed in Chapter 5.

Terry (1994) also explored the effect of disguising individuals on memory for computergenerated faces. Experiments included the addition or removal of glasses, or the addition
or removal of a beard on computer generated face images. Removing glasses reduced
recognition accuracy, as did the addition of a beard, and also the removal of a beard, but
to a lesser extent. However, adding glasses did not affect recognition accuracy. Righi,
Peissig & Tarr (2012) replicated this result with real face images from the TarrLab
database, showing that the addition of glasses did not affect face recognition memory
performance as much as the addition of a wig or the removal of glasses. These findings
show that a change in appearance in form of a disguise is not always clear cut in terms of
the effect it will have on face recognition – some manipulations impair performance
whereas other do not.

Both of these studies (Patterson & Baddeley 1977 & Terry, 1994) report overall
impairments to face recognition memory when an image is presented in a changed form
(in a disguise) at test compared to at presentation despite significant difference in the
disguise stimuli used in each experiment. Patterson & Baddeley’s (1977) Experiment 2
and Righi et al. (2012) consisted of prescribed disguise stimuli whereas Terry (1994) used
computer-generated stimuli images. This prominent effect of disguise on face memory
makes it seem likely that face-matching tasks will also be affected by disguise
manipulations as factors such as changes in pose and expression, which have influenced
face recognition memory, have also reduced performance for face-matching in past
studies (Bruce, 1982; Hancock et al. 2000).

147

Disguise in Face-matching
Dhamecha, Singh, Vatsa & Kumar (2014) are the only researchers to have tested human
face-matching accuracy for disguised faces. Their study compares human face recognition
performance to machine algorithm performance. The main reason for human
performance being investigated was to learn from the strategies used by humans in order
to enhance the algorithms.

Figure 5.1 Images from the IDV1 database. Props include glasses, fake beards and moustaches, medical
maskes and hats turbans.

Dhamecha et al. (2014) created the IIII-Delhi Disguise Version 1 face database (IDV1) for
their study (see Figure 5.1). This consisted of images of 75 models each of whom had
been given access to accessories that they could use to disguise themselves. Models were
photographed in 5 disguises, and in 1 no disguise (neutral) condition. All photographs
were taken under nearly identical lighting conditions, with a frontal pose and neutral
expression, thereby limiting options for creating disguise. Accessories for disguise were
wigs, fake beards and moustaches, sunglasses and glasses, hats, turbans, veils, and
medical masks. Models could wear just one, or multiple accessories for their disguises but
were simply told ‘to use the accessories at their will in order to get disguised’, therefore
the models conducted simple free disguise. The props and the study itself focus towards
occluding specific areas of the face (e.g. eyes hidden by glasses, medical masks hiding
mouth). Occlusion of features is a typical result of many simple disguise manipulations.
The authors point out that the photographs used as stimuli included all parts of the face
being hidden at least once. Although occlusion of features may be effective in hiding
identity, it is important to remember that concealing parts of a face would often not be
an effective disguise in the context of identity fraud as these disguise props would have to
be removed for facial comparison inspections. Concealment of features through use of

148

props is a different type of disguise to manipulating the way features themselves look and
may lead to very different results.

The study reported that human performance accuracy for matching the face pairs
differed as a result of ethnicity of the faces viewed and also familiarity with the faces.
Performance accuracy was lowest for matching unfamiliar faces of different ethnicity
(66% accuracy), followed by unfamiliar faces of the same ethnicity (69.5%) and highest for
the familiar faces, all of which were the same ethnicity (75% accuracy). Performance of
the algorithm proposed in the paper was comparable with the unfamiliar, different
ethnicity human viewers (Dhamecha et al., 2014).

Dhamecha et al. (2014) provide the first assessment of disguise face-matching accuracy
by human viewers. In particular, the free disguise aspect used by Dhamecha and
colleagues was interesting in that it produced a stimuli set of faces that differed from
each other in their disguises, with multiple disguises for each face. However, this design
also had limitations. One specific area of concern is that there were inconsistencies in the
matching task. Participants in the study made same or different identity judgments to
pairs of face images. Sometimes this involved matching a neutral face to a disguise face,
but more often participants were matching pairs of disguised images. Nearly all the faces
had some type of disguise present, therefore the study compared performance between
familiarity and ethnic groups on a task of matching disguised faces rather than
investigating whether disguise impaired performance more than the matching of
undisguised face pairs. This design meant that a direct comparison between performance
accuracy for disguised and undisguised faces could not be made. An additional limitation
of the design is that images were cropped so that only internal cues could be used. This
disregards any influence of external cues which may actually have been able to help
contribute, or alternatively could have worsened, human participants’ face-matching
performance. This would have been an interesting comparison to have, particularly as
previous work suggests that unfamiliar viewers rely heavily on external features when
making facial identity judgments (see Chapter 6 for further discussion of features and

149

identity judgments) (Bruce et al, 1999, 2001; Bonner, Burton & Bruce, 2003; Megreya &
Bindemann, 2009).

Limitations of Previous Disguise Research
There are four main limitations of the previous research conducted on facial identity
performance for disguised faces. These are i) that past tasks compare recognition of
disguised faces with identical image comparisons, ii) a focus on memory tasks, iii)
investigation of only evasion disguise, ignoring impersonation, and iv) the use of simple,
often prescribed, disguises as the stimuli. The reasons why each of these points limit
disguise research is discussed below.

Firstly, performance for disguised face identity judgments have most often been
compared with performance for remembering exact images. Patterson & Baddeley,
(1977) forced comparisons of image recognition (remembering the exact image) with a
difficult case of face recognition (different images of the same face), rather than testing
whether disguised face recognition is more difficult than undisguised face recognition
(across different undisguised images of the same identity), which would be a more
meaningful comparison. Some of the studies have actually made alterations to the
learning image itself to create a disguise condition, which recreated a scenario more
similar to image matching than purely investigating disguise (e.g. Terry 1994).

Recognition of identical images is an easy task for humans (Bruce, 1982). Facial
appearance changes incidentally across multiple different images. In real life facematching scenarios, it is performance across these types of changes that are of interest,
as it not possible to capture the exact same image of a person across time. This has now
been recognised in the study of face recognition. Ambient face images, which include
naturalistic and incidental changes, such as differences in expression, are now typically
the image type used in tasks that try to capture facial identity performance accuracy.

150

Comparisons with image matching performance do not match on to the real world
problem where naturalistic differences exist between multiple images of a face. In the
investigation of disguise, it is thus important that performance for disguised faces is
compared against performance for faces that include this incidental change in
appearance

Further to the focus on image matching, past research is mostly confined to the study of
face recognition memory for disguised faces (Patterson & Baddeley 1977; Terry, 1994;
Righi et al. 2012). This thesis focuses on face-matching accuracy, as this is a frequently
used identification check that does not rely on any memory component. The only task to
date which has looked at face-matching accuracy for disguise faces (Dhamecha et al.
2014) investigated the performance of computer algorithms on the task compared to
human performance, but did not provide a control of performance for faces in no
disguise. Previous research has found that unfamiliar face-matching is poor in cooperative
facial image comparison task, but I believe it to be interesting to know whether disguise
impairs face recognition performance further. In criminal or undercover police situations
there can be very strong incentives to carry out a realistic and successful disguise, but it is
not yet confirmed that disguises impair face-matching performance.

Previous research on disguise has focused exclusively on evasion – changes to appearance
that make the model’s own identity difficult to determine. Disguise could also involve
impersonation, changing appearance to look like a specific other person, but this has not
been addressed by disguise research to date. Most stimuli databases include exclusively
male faces and some studies have relied on computer generated stimuli rather than real
human face images to explore evasion disguise. A database of male and female human
faces in both evasion and impersonation disguise is necessary to better understand
disguise.

151

A final limitation of the previous research on disguise is that the stimuli images used have
been unsophisticated. Simple disguises have dominated which involve the addition or
removal of props to a face from a limited supply of props provided by the experimenters.
This limits the ways in which individuals can disguise themselves and may not reflect the
props that the person would choose to disguise themselves naturally. Simple disguise has
often led to occlusion of features. This disguise technique would create ineffective
disguises in many contexts of identity fraud. Most disguises have been prescribed by the
experiments, meaning that no information has been gathered on the changes that people
naturally make to their appearance to create disguises. Patterson & Baddeley
(Experiment 1) were the only researchers to test more naturalistic disguises in terms of
the actor photographs, but the intent at time of photograph for these appearance
alterations was not specifically disguise. Examples of existing disguise face databases are
shown below - figures make limitations of the stimuli involved evident.

5.3 Existing Disguise Face Databases
The AR & IDV1 Databases
The AR database (Martinez & Benavente, 1998) and the IDV1 database (Dhamecha et al.,
2014) are the only existing disguise face databases to contain images of real people under
a variety of disguises. In the case of the AR database, disguises include changes in lighting
and deliberate changes in expression, which led to incidental changes in the appearance
of the identity, and also more deliberate changes to appearance through the addition of
props – sunglasses and a scarf to hide features of the face (Figure 5.2). This database
consists of only changes that constitute evasion disguise and incidental appearance
change. The IDV1 database (image shown previously in Figure 5.1) kept expression and
lighting constant; therefore disguise was created purely through the addition of props.
Both the AR database and the IDV1 database show rather unconvincing disguises, in that
it is obvious that people are trying to hide their appearance through props.

152

Figure 5.2 Images from the AR database. Disguise manipulations are limited to a change of expression or
the addition of sunglasses or a scarf.

The National Geographic Database

The National Geographic Database (Ramathan, Chowdhury & Chepella (2004) also
includes real faces image with a disguise component, but consists of 46 images of just one
identity. It is questionable whether this database should really be considered a disguise
database at all, as many of the photographs (e.g. the top row in Figure 5.3) fit the
category of incidental change as they show natural variation between multiple images of
a face rather than a deliberate attempt to change appearance. The bottom row (Figure
5.3) does consitute disguise images, but these disguises are very obviously fake. It
appears that the disguises have been applied to the images themselves (e.g. addition of
beard or moustache) rather than to the model prior to the photograph being taken.

153

Figure 5.3 A sample of images from the National Geographic Database (Ramathan, et al., 2004).

TarrLab Face Database
A final example of human disguise face images used in the previous research is the
TarrLab face database. This includes images of faces photographed naturally and also the
same identities photographed wearing a wig and glasses (Figure 5.4). Sometimes these
glasses are reading glasses and at other times they are sunglasses, which occlude the
model’s eyes. An advantage of the TarrLab database is that images are of both males and
female faces, whereas all other databases discussed consist of male faces only.
Additionally, the fact that models are photographed both with and without disguise
allows direct comparisons of disguise to be made. The database is however limited to
prescribed simple disguise, and many of the disguises include occlusion of features. As
discussed previously, items that occlude features may have to be removed in security
screening scenarios.

154

Figure 5.4 Example images taken from the TarrLab face database.

Synthetic Face Database
Disguise face databases made up of images of human faces have been rather limited to
date. Singh, Vatsa & Noore (2009) acknowledge and address this by creating a new
disguise face database which encompass a greater range of disguises, but are not images
of real human faces (Figure 5.5). Singh et al. (2009) used Faces Software, which is a
programme used to make face images based on facial descriptions in police
investigations, to build synthetic faces in different disguise conditions. Computer
algorithm performance was then tested on this database, with the best performing
algorithm performing with 71% accuracy when dealing with images that had multiple
components of disguise. The study looked exclusively at algorithm performance on the
disguise face database, with the aim of identifying which algorithm performed with
highest accuracy on this image set. There was no control condition to compare
performance for undisguised face images. This study is therefore limited to evasion
disguise performance for computer algorithms on a set of computer generated disguised
face images. Moreover the disguises are not realistic, for example, a pirate style eye patch
is unlikely to be worn by a disguise perpetrator.

155

Figure 5.5 Examples from the synthetic face disguise database (Singh et al. 2009).

5.4 FAÇADE Database
Shortcomings of Existing Databases
As seen from previous research, existing databases are limited - they are unrealistic or
involve occlusion of features and include only evasion disguise images. It is also extremely
important to investigate impersonation disguise, as in some instances people may have
very strong incentives to pass themselves off as a specific other person. Impersonation
could include the case of illegally accessing a country on a stolen passport.

Furthermore, past databases are rather unsophisticated in the disguise manipulations
applied. All of the disguise databases rely on simple disguise. Many are prescribed
disguises (whereby the experimenters have told the model’s exactly how to apply the
props) but in cases where models have been free to disguise themselves as they wished
disguises were severely restricted to the addition of a limited selection of props and
combinations of these. As models were never given the opportunity to disguise

156

themselves with free reign over changes to their appearance, very little is known about
the disguise process itself, let alone, what works.

With these shortcomings in mind, I created my own database of disguised and
undisguised faces, named the FAÇADE image database, which I will use to investigate
deliberate disguise.

Goals for the FAÇADE Database
I created the FAÇADE image database to allow face recognition to be explored more fully.
Creating a disguise database from scratch allowed the effect of disguise manipulations
and the process in creating disguises to be investigated in more detail than ever before. A
major goal of this database was to allow research to be conducted to directly compare
face-matching performance for disguised against undisguised faces. In addition to this I
wanted to break down disguise further – to explore evasion disguise and compare and
contrast this to impersonation disguise for the first time.

Unlike past disguise databases, which were made up of images of models who had
applied prescribed simple disguises, I wanted to be able to explore how people disguised
themselves when they were given the opportunity to create their own disguises. Use of
free disguise will allow insight into what measures people would most naturally take to
disguise themselves and also allow me to explore what disguises in the database worked
best based on results from the matching task. This exploratory analysis will be conducted
in Chapter 6.

My initial goal and focus, following the theme of face-matching performance in
challenging situations investigated throughout the rest of this thesis, was to assess
whether face-matching performance is influenced by the presence of disguise.

157

Approach to the Image Acquisition
As mentioned above, I distinguish between two different types of disguise within my
disguise stimuli—Evasion (trying not to look like oneself) and Impersonation (trying to
look like a specific target person). To capture this distinction, I gave 26 volunteer models
(i) a photo of themselves, and asked them to make themselves look unlike that reference
photo for a subsequent shoot (Evasion condition), and (ii) a photo of someone else, and
asked them to make themselves look like that person (Impersonation condition). It is
possible that similarity of the impersonation face may influence disguise effectiveness. To
investigate this, two impersonation conditions were present for each identity –
impersonating a similar face (rated to be the most similar out of 33 match faces by a
group of 3 viewers) and impersonating a face that was selected at random. A no disguise
photograph was also taken of each model.

Models were instructed that the task was to look as much like or unlike the identity of the
target profile photograph, rather than the image itself i.e. I was interested in an identity
match task rather than image matching, which itself has already been addressed as a
shortcoming of previous research.

One of my goals for the FAÇADE database was to create a database that would not only
consist of realistic disguise stimuli for both undisguised, evasion disguise and
impersonation disguise faces, but in the acquisition of which, could further the
understanding of the things which people do to create each of these disguises. Models
were thereby unguided in their disguise effort, as it was of interest to us to find out what
people did to disguise themselves. To allow this exploration of changes made, models
wrote down all changes they made to their appearance, and the intent of each change, as
they carried out their disguise transformations.

158

I provided models with any props that they wanted to aid their disguises. Props were
ordered on request to match the models wishes, but including clothing, wigs, hair
accessories, makeup, glasses and jewellery (see Figure 5.6 for selection of props used).
Hats, sunglasses and any other occluding items were not allowed, as these props have to
be removed if someone was physically present in these items at passport security.
Models were instructed that their resulting appearance must not be considered out of
place as being a real person’s work identity badge image, rather than fancy dress.

Figure 5.6 Sample of props used to create the disguise face database.

Photographs were taken at a time that suited the model, with photographs (evasion,
impersonation similar, impersonation random) not necessarily captured on the same day.
This was left to personal choice of the model and gave our models time to make changes
to their appearance between photography sessions such as waiting for hair and beard
growth.

Motivation
It was important that our models were highly motivated and dedicated to create
convincing disguises. This was a very time consuming task for our models and required

159

much thought and some made large changes to their appearance including having their
hair cut or died, or beard shaven/grown. It was up to the models which changes they
made, so they could decide how far they wanted to go with the disguises. Models were
motivated by cash reward, with the best disguise (based on performance in Experiment 1)
in each condition (evasion, impersonation similar, impersonation random) receiving a
cash reward of £50. The models were extremely dedicated to the task and became
competitive with one another to create the best disguise.

Selecting the best disguise images
Several images (9 images) of the models were taken in each disguise condition, with
models varying aspects such as facial expression, pose and lighting to try and create the
best image for each disguise condition (see Figure 5.7). A group of four unfamiliar viewers
(the stimuli selection group) worked together to decide upon the most convincing match
or mismatch images for each model in each condition. The stimuli selection group viewed
the Impersonation (similar and random) images and the Evasion images for each of the 26
models alongside the corresponding reference image of the target face (the model’s own
face in evasion condition, and the faces of the people the models were trying to
impersonate in the impersonation conditions). The group knew that they were dealing
with images of people in disguise, and were informed of the true identity classification
(same or different) in each matching situation to aid them in their decisions.

160

Figure 5.7 Image taken during stimuli selection process.

In order to better understand what was driving the decisions made by the stimuli
selection group I asked them the following questions for each identity in each disguise
condition: 1). Which image provides the best match? 2). How good is the disguise (rated
on a scale of 1 – 7 with 1 representing ‘not a good disguise at all’ and 7 indicating an
extremely good disguise)? 3). What is it about the chosen image which makes it an
effective disguise? The results of each of these are discussed in Chapter 6, Understanding
Disguise, but for now I will continue to focus on the database itself and the image pairs
that will be used to test face-matching performance.

Face Image Pairs
The images chosen by the stimuli selection group provided the final disguise image for
each of the models in each of the disguise conditions. All other images were disregarded,
with only the most convincing disguises for each model kept in the FAÇADE database to
satisfy the disguise face image conditions. I was then able to create face image pairs for
both disguise and no-disguise conditions using the database images. The face-matching
task included the following face image pairing conditions: same identity no disguise, same

161

identity evasion disguise, different identity similar no disguise, different identity similar
impersonation disguise, different identity random no disguise and different identity
random impersonation disguise.

Figure 5.8 Example pairs for each condition. Top row shows same identity pairs, the lower two rows show
different identity pairs. Pairs in the first column are in no disguise. Pairs in the second column are in
disguise. All 26 models were photographed in each of the conditions.

Examples of the stimuli pairs are shown in Figure 5.8. The image pairs on the left show no
disguise image pairings. The top two rows in this column consist of pair types that are
very similar to those seen in other standardised face-matching tasks such as the GFMT.
Pairs of images can show either the same person or two different people. I have extended
the design, through introducing the third column to include two types of different person
trial. Similar pairing means that the foil here is the most similar looking person from the
option of face images. Random pairing means that the foil is drawn at random from

162

images of models of the same sex. Therefore, random pairings can be rather dissimilar in
appearance, showing two faces side by side, for which one would not normally expect the
identities in the pair to be confused.

This led to a very convincing (and rather confusing) series of images within the FAÇADE
database (see Figure 5.9).

Figure 5.9 Selection of images taken from the disguise base database to create a wheel of disguise. Images
with the same colour frame show the same identity. Images with different colour frames are of different
identities.

Familiarity
Familiarity has been an overarching theme throughout this thesis. It has been repeatedly
shown that people who are familiar with a face are better at matching images of that face
than people who are unfamiliar with the face. Now the familiarity advantage is being
pushed even further than before as I test whether familiarity with a face can help in the

163

case of deliberate disguise. In theoretical terms, as a face becomes familiar, the normally
accepted range of variability for a face is being learnt. Effective disguise involves taking a
face outwith that normal range of accepted appearance. It is thus possible that in the
case of disguise, familiarity with the face or faces involved may not help much, precisely
because disguising a face takes it out of the accepted range of appearances and hence
out of the familiar viewer’s area of expertise.

Dhamecha et al. (2014) argue that familiarity with a face does help with face-matching
performance for disguised faces. In their experiment Dhamecha and colleagues (2014)
define familiar viewers as participants who worked in the same department as the
models whose images were used in the matching task, and unfamiliar participants as
people who did not have previous encounters with the models. Although there is a
significant effect of familiarity when the results of trials are pooled, the effect is not
significant for both trial types considered in isolation. For same person trials, familiar
viewers were more accurate at the task then unfamiliar viewers. However for different
person trials, incorrect matches were made equally often by familiar and unfamiliar
viewers. Both items in both match and mismatch pair items were most often disguised.
Based on these findings, the familiarity advantage seems uncertain for the case of
disguise.

I expect that unfamiliar viewers will make more mistakes on the disguised face trials than
familiar viewers will, as their representation of these faces are less finely tuned than that
of familiar viewers. Unfamiliar viewers frequently accept imposter faces to be a target
face. This is demonstrated by both earlier work in this thesis (e.g. Chapter 2) and the
results of face-matching tasks such as the GFMT, which show around 80-90%
performance accuracy for unfamiliar viewers, confirming that people are making errors in
identity judgment.

164

There are however instances of extremely familiar viewers being tricked by imposter
scenarios. Thompson, (1986) demonstrated that it is possible to mistakenly reject a family
member if context interferers with the identity judgment. Thompson, (1986) constructed
a scenario where a daughter walked past her parents, without acknowledging them, while
they were on holiday. The parents believed that their daughter was at home, in a
different location. The parents concluded that the identity was not that of their daughter,
based on the incorrect contextual information that they held. This example demonstrates
that familiarity does not guarantee correct identity judgment. It is possible that disguising
a face may have a similar effect, as some context may be lost when a face is disguised
with the attempt of removing a face outwith its accepted range of appearances. For
example, a person may normally be seen in work clothes with a professional and
approachable appearance, but in disguise their displayed persona may be different.

The Study – Research Questions & Predictions
This study will use image pairs from the FAÇADE database to answer four main research
questions. Firstly, I will address whether overall face-matching performance accuracy is
worse for disguised than undisguised face pairs. Presumably it will be more difficult to
match disguised faces than images of cooperative stimuli.

The second question of interest is whether all disguise manipulations cause equal
impairment to face-matching performance or if disguise type has an effect. Results of
past experiments show that more errors are made for same person trials than different
person trials in face-matching tasks. This demonstrates difficulties in integrating multiple
images of the same identity (e.g. Jenkins et al. 2011). Thus, a deliberate attempt to
frustrate the integration of identity (the evasion condition) will likely make the task of
‘telling the faces together’ harder still. Therefore I predict that trials involving evasion
disguise will be more error prone than impersonation disguise trials.

165

I will also investigate whether there is an effect of impersonation type, i.e. is it easier for
someone to pass themselves off as being a person who looks naturally similar in
appearance to them or are they equally effective at looking like somebody who is
selected at random? Past research suggests that impersonation similar disguise will cause
more errors than impersonation random disguise. Light, Kayra-Stuart & Hollander (1979)
found that people were worse at remembering faces that looked similar to a prototype
face than those that looked unusual in appearance, suggesting a role of distinctiveness. I
thereby expect that impersonation random disguises will lead to fewer errors in the facematching task than impersonation similar disguises.

Each of these questions will also be addressed with regard to familiarity with the faces
viewed. As discussed in the familiarity section, based on theories of face learning,
familiarity is not guaranteed to help in case of disguise. Familiar viewers will however be
used to viewing the models’ faces over a wider range of appearances than our unfamiliar
viewers. Therefore familiar viewers are predicted to be better than unfamiliar viewers at
the face-matching task, although familiar viewers may also be somewhat affected by the
disguise manipulations. In total I will conduct three disguise face-matching experiments,
each of which manipulate viewer group but will also address each of the research
questions here outlined. Experiment 10 will test face-matching performance of unfamiliar
viewers, Experiment 11 will test performance of unfamiliar viewers who are informed of
the disguise manipulation and finally Experiment 12 will access face-matching
performance for familiar viewers.

5.5 Experiment 10: Unfamiliar Viewers
Experiment 10 examines face-matching performance for unfamiliar viewers on a task
involving disguised and undisguised image pairs. The experiment will use the images from
the disguise face database, therefore allowing an investigation of face-matching
performance accuracy for evasion, impersonation similar and impersonation random
disguise.

166

I predict that disguise will impair face-matching performance, with evasion disguise
making face pairs most difficult to correctly identify, and impersonation disguise will also
cause difficulties but to a lesser degree. I believe impersonation similar disguises will be
easier to execute than impersonation random disguises, hence more errors will be made
when matching impersonation similar than impersonation random faces.

Method
Participants
30 undergraduate students (M = 8, mean age = 23) from the University of York
volunteered as participants in this study in return for payment or course credits.

Design & Stimuli
The experiment took the form of a 2x3 within subject design, with factors Disguise
Condition (levels: Disguise and No Disguise) and Pair Type (levels: same, different similar
and different random). The dependent variable was performance accuracy in the facematching task for each of the independent variables listed above.

Stimuli were the face image pairs from the FAÇADE database. Stimuli were kept constant
across all experiments in this chapter.

Procedure
Participants viewed the image pairs from the FAÇADE face database on a computer
screen. The viewing distance was 50cm from the screen. Image pairs were presented one
at a time as part of a self paced face-matching task. The participants’ task was to decide,
for each image pair, whether the two images were of the same person’s face or were

167

images of two different people – pressing ‘z’ on the keyboard for same pairs, and ‘m’ for
different. All image pairs were presented in a randomized order and over the whole
experiment the participants saw each of the models in each of the disguise conditions. In
total participants viewed 156 image pairs.

It was important to check that all items in the task were unfamiliar to the participants in
this experiment. Following the face-matching task, participants were given the reference
images which they had viewed in the task, and asked to indicate how familiar they were
with the person’s face before the experiment. Familiarity was measured using the
familiarity scale that I designed in Chapter 1 of this thesis. Participants placed the face
cards on the familiarity scale that ran across the desk, indicating familiarity on the scale
from 0 (completely unfamiliar) to 100 (extremely familiar).

Result
Participants’ responses from the face-matching task were broken down into each of the
Disguise Conditions and Pair Types, with a mean score of performance accuracy
calculated for each. These means were then combined across all participants and
averaged to calculate the overall mean for each condition. These breakdowns of scores by
disguise condition and disguise type allowed the research questions of whether there was
an overall effect of disguise on face recognition accuracy, and whether disguise type
influence accuracy, to be answered.

168

Performance Accuracy for Unfamiliar Viewers
100

Accuracy (%)

90

80

No Disguise
Disguise

70

60

50
Same

Different Similar

Different Random

Pair Type

Figure 5.10 Performance accuracy for unfamiliar viewers for evasion, impersonation similar and
impersonation random pairs when the images consisted of no disguise or disguise pairs. Error bars show the
standard error of the mean.

A within subjects ANOVA with factors of Disguise Condition and Pair Type was conducted
to find out whether disguise presence affected face-matching performance and whether
disguise Pair Type affected matching accuracy (see Figure 5.10). As predicted, there was a
significant main effect of Disguise Condition. Participants performed more poorly for
Disguise (M = 77.39 % accuracy, SD = 21.27) than No Disguise face pairs (M = 95.26%
accuracy, SD = 5.97) [F(1, 29) = 75.88, p<.001, η p2 = .72]. There was also a significant main
effect of Pair Type (F(2,29) = 22.87, p<.001, η p2 = .44] and a significant interaction
between Disguise Condition and Pair Type [F(2,29) = 37.95, p<.001, η p2 =,57]. This showed
that both the presence of disguise and also the type of trial affected performance
accuracy rates, but accuracy was not influenced in the same way by Pair Type (Same,
Different similar, or Different random) in each of the disguise conditions.

169

Simple main effects were conducted to find out where the significant difference lay. Pair
Type was significant for Disguised [F(2,116) = 53.79, p<.001, η p2 = .48] but not for No
Disguise pairs [F(2,116) = 2.03, p>.05, η p2 =.03], demonstrating that participants showed a
difference in their accuracy of matching across different pair types only for the disguise
present faces. Disguise presence significantly affected performance for each of the Pair
Types; Same identity pairs [F(1,87) = 151.5, p<.001, η p2 = .64], Different (Similar
appearance) identity pairs [F(1,87) = 12.78, p<.005, η p2 = .13] and Different (Randomly
matched) pairs [F(1,87)= 9.19, p<.005, η p2 = .1].

Tukey tests revealed that there were significant differences between each level of trial
type for the disguise present faces. Same identity pairs (represented on the graph as
Same, condition Disguise) caused most errors (M = 60.38% accuracy, SD = 25.69), then
Different Similar identity pairings (those on the graph shown as Different, and condition
Disguise) (M = 82.18% accuracy, SD = 12.8), and finally Different Random identity pairings
(M = 89.62% accuracy, SD = 9.32).

Discussion
This experiment found that disguise did significantly impair face-matching performance
for unfamiliar viewers. Therefore, in relation to my first research question, and in
accordance with the hypothesis, matching disguised faces is more difficult than matching
faces that are not in disguise. My second research question, which asked whether all
disguise manipulations cause equal impairments to face recognition, is also addressed by
this experiment. The results were again as predicted, with evasion disguise affecting facematching accuracy more than impersonation disguise. With regards to research question
three, there was also a significant effect within impersonation itself. Unfamiliar viewers
made more matching errors when the Impersonations involved someone who was of a
similar appearance to the model than when the impersonation pairings were at random.

170

The results from this experiment will later help to answer research question four,
regarding matching accuracy and familiarity, providing data from unfamiliar viewers
which can be compared with that of familiar viewers when collected.

These findings using the FAÇADE database address the issue of disguise face-matching in
more detail than any previous study. However, I want to continue to understand disguise
even further. I found that face-matching performance was poor for disguised faces
compared to faces that were not in disguise, and although difficult to compare directly
with other studies, performance for the disguised faces in my task was poorer than
performance on the full version of the GFMT - mean performance accuracy was 77% for
the disguised face pairs, whereas performance on the full version of the GFMT is 87%.
Performance may have been particularly poor in my experiment because viewers were
unaware that faces they viewed may have been disguised. It is possible that knowledge of
the possibility of disguise presence could improve face-matching performance.

5.6 Experiment 11: Unfamiliar (Informed) Viewers
In Experiment 10 unfamiliar participants’ face-matching performance was impaired by
disguise presence. Making accurate identity matching judgments for the disguised faces
was a difficult task, but it is possible that this task was difficult because participants were
not expecting to see faces that were disguised. Expectations have been reported to
influence performance accuracy in visual search and other cognitive tasks (e.g Alain &
Proteau, 1980; Kastner, Pinsk, De Weerd, Desimone & Ungerleider, 1999). If participants
are told about the disguise aspect of the study then performance may improve. This will
now be tested in Experiment 12.

If poor performance from Experiment 10 is because participants did not know to look for
disguises, their performance accuracy should be far better in Experiment 11 than it was in
Experiment 10. If performance accuracy for disguised faces is no better when participants

171

have this knowledge, then being naïve to the disguise component is not what makes the
task difficult.

The research questions remain the same as in Experiment 10, but this time I investigate
whether the results of Experiment 10 remain when participants are informed of disguise.

Method
Participants
Thirty undergraduate students from the University of York who had not taken part in
previous experiments involving the FAÇADE database, volunteered as participants for this
experiment (mean age = 21, M = 11).

Design & Stimuli
The experiment took the form of a 2x3 within-subject design, with factors Disguise
Condition (levels: Disguise and No Disguise) and Pair Type (levels: Same, Different Similar
and Different Random). The dependent variable was performance accuracy in the facematching task for each of the independent variables listed above. This will enable the
research questions outlined in the introduction to be addressed.

To test whether performance differs for disguised faces between Experiment 10 and
Experiment 11, a between subjects design will be used.

Stimuli were the face image pairs from the FAÇADE database. Stimuli are kept constant
across all experiments in this chapter.

172

Procedure
Experiment 11 was carried out exactly like Experiment 10, the only difference to the
method being that this time participants were made aware before beginning the
experiment that some of the face images they view may be disguised to either look unlike
themselves or to look like another person. It was clearly stated that the face-matching
decision always concerned an identity judgment based on the true identity of the face.
Participants were to decide if the images they saw were really images of the same person,
or if they were images of two different people.

Results
Results were cacluated exactly as in Experiment 10, with mean performance accuracy
broken down by Disguise Condtion and Pair Type. In order to answer the research
questions regarding an overall effect of disguise (Disguise Condition), possible differences
between Pair Type, a within subjects ANOVA was conducted.

173

Figure 5.11 Performance accuracy of unfamiliar viewers who were aware of the disguise component of the
face-matching task. Error bars show the standard error of the mean.

A significant main effect was found for Disguise Condition - participants were worse at
matching Disguised faces (M = 80.51, SD = 18.68) than faces with No Disguise (M =
95.30% accuracy, SD = 6.62) [F(1,29)= 44.25, p<.001, η p2 = .6]. There was also a significant
difference main effect of Pair Type [F(2,29)= 21.39, <.001, η p2 = .42] and a significant
interaction between Disguise Condition and Pair Type [F(2,29) = 55.52, <.001, η p2 = .66]
(see Figure 5.11).

Simple main effects highlighted there was a significant effect of Disguise Condition on
performance for each of the Pair Types (same [F(1,87) = 119.56, p<.001, η p2 = .58],
Different Similar [F(1,87) = 9.45, p<.005, η p2 =.1] and Different Random [F(1,87) = 11.74,
p<.001, η p2 = .12]). There was also a significant main effect of Pair Type for both Disguised

174

[F(2,116) = 56.21, p<.001, η p2 = .49] and No Disguise faces [F(2,116) = 3.64, p<.01, η p2 =
.06].

Tukey tests revealed that there was a significant difference in performance between each
level of Pair Type for the Disguise faces. Same identity pairs (represented on the graph as
Same, condition Disguise) caused most errors (M = 67.60% accuracy, SD = 20.01),
followed by Different Similar identity pairings (those on the graph shown as Different, and
condition Disguise) (M = 84.49% accuracy, SD = 15.73), and finally Different Random
identity pairings (M = 89.36% accuracy, SD = 10.09). For No Disguise faces the only
significant differences lay between performance accuracy for Different Similar (M =
92.31% accuracy, SD = 8.86) and Different Random face pairs (M = 98.08% accuracy, SD =
3.46)

In order to test whether performance is better if participants know of the possibility of
disguise manipulations than when they are naïve to the disguise component, I carried out
a between subjects ANOVA to compare performance accuracy for disguised faces in
Experiment 10 with that of Experiment 11. Interestingly participants in Experiment 11,
who knew that some of the faces in the experiment were disguised to not look like
themselves or to look like a specific other person, were no better at the face-matching
task with this knowledge of the disguises, than participants in Experiment 10 who were
not informed of the disguise manipulations before the experiment F(1,58) = .625, p = .43,

η p2 =.01. Thus, knowing to look out for disguise does not help to see through disguise.

Discussion
The same pattern of results was seen for each of the research questions for Experiment
11 (when participants were aware of the disguise element) as for Experiment 10 (when
participants were not aware that faces could be in disguise). Experiment 11 found that
informed viewers were significantly worse at matching disguise face pairs than faces that

175

were not in disguise, evasion caused more difficulty than impersonation disguise, and
impersonation similar trials were harder than impersonation random pairs.

The unique research question for Experiment 11 was whether performance accuracy for
disguised faces would be better than it was in Experiment 10. No significant difference
was found in performance for matching disguised face pairs in each of these experiments,
therefore knowledge of disguise does not make the task easier. Alternatively, it is possible
that participants in Experiment 10 may have figured out the disguise manipulation for
themselves whilst doing the task, either way performance remains poor for disguised
faces in both the experiments, suggesting that matching disguised faces is an extremely
challenging task regardless of holding knowledge that disguises may be present amongst
the image pairs.

5.7 Experiment 12: Familiar Viewers
There is a huge distinction in the face recognition literature between face-matching
performance for familiar, compared with unfamiliar viewers. Familiar viewers perform
with impressive levels of accuracy in matching tasks even when the images available for
comparison are of poor quality (Burton et al. 1999). Chapters 2 and 4 show further
situations where familiarity improved performance in challenging face-matching tasks. I
now test familiarity on perhaps our most challenging task to date – the situation of
deliberately disguised faces. Can familiarity with a face help us see through deliberate
disguise and hence make the viewer immune to deception of disguise?

Jenkins et al. (2011) highlight that familiar viewers are able to easily see through natural
changes in a face. The distinction between familiar and unfamiliar face recognition
performance may be due to experience of variability in photos of the same face. They
argue that when someone is becoming familiar with a face they are learning all of the
different appearances, which that face can take. Any face that falls within the excepted

176

range of faces for a specific person will be considered a match for that identity, any image
which does not fall within the range of accepted appearances for a face will not be
considered a match. It is believed that the range of accepted faces for a given face is
smaller for unfamiliar viewers and larger for familiar viewers who have had more
experience with the face. Therefore the idea is that the acceptable range of facial
appearances for a given person becomes more finely tuned as their face is learned. This
explains why familiar viewers can see through changes in viewpoint, expression and
lighting, whereby unfamiliar viewers can be impaired by trivial image changes - they do
not have enough experience with the face to accurately know what appearance the face
would take under these changes (Bruce 1982; Adini, Mosses & Ullman, 1997). However,
this theory does not necessarily cover the situation of disguise.

Presuming that the viewer has not had prior exposure to an individual in their disguised
form, it is unknown whether the previous exposure they have had with the face would be
sufficient to also help the viewer see through a disguised version of the face. Essentially
this is the challenge for creating a convincing disguise. For an evasion disguise to fool a
familiar viewer it must change the appearance of the model so that the new appearance
falls out with the accepted range of appearances held by the viewer for that model.
Impersonation disguises must move outwith the accepted range of the model and into
the accepted range of the reference photograph in order to cause a familiar viewer to
make face-matching errors. If the models have been effective in doing this, familiarity
with a face may not help a great deal, as the face would be taken outwith the familiar
viewers area of expertise.

There are essentially two strands of research questions regarding familiarity. The first
includes the familiar participants alone, and asks whether familiar viewers are able to see
through disguise manipulations. This relates to the first three research questions laid out
in the introduction. If familiarity allows participants to see through disguise, then there
would be no difference in performance accuracy for disguised and no disguise face pairs.
If however, familiar viewers are worse at matching disguised face pairs than undisguised

177

face pairs, then even familiar viewers are affected by an overall disguise manipulation. It
is again possible that some types of disguise may affect performance more than others.
For familiar viewers I predict that impersonation disguise will not be as effective as
evasion disguise. To identify faces in impersonation disguises, participants can use both of
the faces in the pair to look for identity cues, as they are familiar with both of the faces.
Therefore a model’s task is harder as they have to look both unlike themselves and
convincingly like the person they are trying to impersonate, whereas evasion involves
moving from only the accepted range for a person’s own appearance, which can be
achieved in many different ways.

The second strand will answer research question number four, of whether familiarity with
a face aids disguised face-matching. This will involve a comparison of results for disguised
face-matching performance across all of the Experiments conducted in this chapter, thus
comparing familiar viewers with unfamiliar viewers. As familiar viewers have more
experience than unfamiliar viewers with the faces in the FAÇADE database, I predict that
participants will be significantly better than the unfamiliar viewers at matching disguise
faces, simply because they have seen the faces in a wider range of appearances, and
therefore may be better able to see through some aspects of the disguises.

Methods
Participants
The participants in this study were 30 individuals who were colleagues and friends of the
disguise models (M = 14, mean age = 27). Participants saw the models on almost a daily
basis; therefore they were of high levels of familiarity with the models’ faces.

178

Design & Stimuli
The experiment took the form of a 2x3 within subjects design, with factors Disguise
Condition (levels: Disguise and No Disguise) and Pair Type (levels: Same, Different Similar
and Different Random). The dependent variable was performance accuracy in the facematching task for each of the independent variables listed above. This will enable the
research questions outlined in the introduction to be addressed.

To test whether face-matching performance for disguised faces differs as a result of
familiarity, a between experiment analysis will be conducted comparing the results of
Experiment 10 and Experiment 11, with those of Experiment 12.

Procedure
Procedure was as in Experiment 10.

Results
A 2x3 ANOVA was conducted to investigate the effect of Disguise Condition and Pair Type
on familiar participants’ face-matching performance.

179

Figure 5.12 Performance accuracy in the face-matching task for viewers who were familiar with the models
whose images featured in the task. Error bars show the standard error of the mean.

A significant main effect of both Disguise Condition (No Disguise or Disguise) [F(2,58) =
20.01, p <.001, η p2 = .41] and Pair Type (Same, Different Similar and Different Random)
[F(1,29) = 24.99, p <.001, η p2 = .46] were observed. This shows that familiar viewers were
worse at matching disguised faces than faces that were not in disguise, and the type of
disguise made a difference to performance accuracy. There was also a significant
interaction between Disguise Condition and Pair Type [F(2, 58) = 22.23, p <.001, η p2 = .43],
meaning that participants were not equally affected by disguise presence for all disguise
pair types (see Figure 5.12).

To understand where disguise caused impairments to face-matching performance simple
main effects were calculated. These revealed that the only significant differences for Pair

180

Type were for the disguised faces [F (2,116) = 40.97, p<.001, η p2 = .41), and not for No
Disguise faces, p>.05. Specifically, Tukey tests revealed that familiar viewers were
significantly worse at matching disguised evasion (Same identity) faces than
impersonation (Different identity) pairs (M = 86.67% accuracy, SD = 13.54), with no
significant difference in performance levels for matching Impersonation Similar (M =
97.69% accuracy, SD = 5.59) compared with Impersonation Random pairs (M = 97.82%
accuracy, SD = 4).

Figure 5.13 Graph showing performance accuracy for Disguise face pairs for each of the 3 Experiments: U
informed (Experiment 10), U uninformed (Experiment 11), Familiar (Experiment 12).

The effect of Familiarity on performance accuracy for disguised faces can be examined by
comparing results across the three experiments (Experiment 10 - Unfamiliar, Experiment
11 - Unfamiliar informed and Experiment 12 - Familiar) (see figure 5.12). Comparing
performance of all experiments on disguised face pair types showed a strong significant
main effect of Experiment, F(2, 87) = 19.15, p<.001, η p2 =.31.

181

Pairwise comparisons revealed that Familiar viewers were significantly better at matching
disguised faces than both Unfamiliar viewers (Experiment 10) and Unfamiliar Informed
viewers (Experiment 11) and reiterate no difference in performance accuracy in matching
disguised faces between Unfamiliar and Unfamiliar Informed viewers. The mean
difference in performance accuracy between Familiar viewers and Unfamiliar viewers
(Experiment 10) was 17%, CI = 10.8 – 22.53, p<.001. And between the Familiar viewers
and Unfamiliar Informed viewers (Experiment 11) this was 13.5%, CI = 7.68 – 19.41,
p<.001. Mean difference in performance between the two unfamiliar groups on the
disguise present matching trials was 3%, this was not a significant difference, p>.05.

Familiarity Score Comparisons
In each of the experiments in this chapter, participants indicated their familiarity with
each of the faces included in the face-matching task after they had completed the facematching task. Scores were based on familiarity prior to the experiment. I compared
mean familiarity score with the face items, for each participant, across the three
experiments using a one-way between subjects ANOVA. The ANOVA showed a significant
difference in familiarity between the groups [F(2,87) =314.43, p<.001, CI = 22.06 – 37.08.
Post hoc tests revealed that the familiar viewers were significantly more familiar with the
faces (M = 75.56 familiarity rating) than both the unfamiliar viewers (M = 6.81 familiarity
rating, mean difference = 68.75, SE = 3.19, CI = 60.94 – 76.55, p<.001) and the unfamiliar
informed viewers (M = 4.69 familiarity rating, mean difference = 70.87, SE = 3.25, CI =
62.92 – 78.81, p<.001). There was no significant difference in the familiarity ratings for
the two unfamiliar groups (mean difference = 2.12, SE = 3.25, CI = -5.82 – 10.06, p>.05).

Discussion
With reference to research question 1, which posed the question of whether disguise
impairs face-matching performance, familiar viewers were significantly worse at matching

182

disguised face pairs than faces that were not in disguise. Whereas previous face-matching
tasks find ceiling performance levels for familiar viewers, in the case of disguised faces,
familiarity can not completely overturn the effect of disguise. Research question 2,
whether all disguise types are equally challenging, is particularly relevant in the case of
familiar viewers. The familiar viewers were worse for disguised faces, only when these
were evasion faces, with no significant difference in performance between disguise and
no disguise recognition for impersonation faces. Question 3 considers differences
between types of impersonation disguise. Unlike in the previous experiments,
impersonation pair type did not matter – familiar viewers could see through both types of
impersonation disguise equally well.

Research question 4, searching for an effect of familiarity, was also answered by this
experiment, in the between experiment analysis. Familiarity improved face-matching
performance for disguised faces compared to that of unfamiliar viewers. Familiar
participants (Experiment 12) were 15% better than unfamiliar participants (averaged
result of Experiment 10 & Experiment 11) at correctly identifying same person and
different person disguise image pairs. Familiar viewers were 11% better than the average
result of the two unfamiliar viewer experiments at matching impersonation faces and
23% better at matching evasion face. These results suggest that when identity judgments
involving disguise need to be made, recruiting familiar viewers to make the identity
judgment would likely lead to a more accurate judgment of identity.

These findings fit with my theoretical interpretations of both familiarity and effective
disguise. It has been argued that when people become familiar with a face what they are
learning is all the different ways that that face can look (Jenkins et al., 2007). For familiar
viewers then, they will be familiar with a far greater range of appearances that any one of
our models’ face can take, than the unfamiliar viewer group will be. This may mean that
some of the changes to appearance, perhaps expression and pose changes for instance,
are learnt appearances of the face in question for the familiar viewers, and hence would
not disguise the face as it may for unfamiliar viewers. Previous exposure of familiar

183

viewers to some of the facial changes used by the models to achieve their disguises, likely
explains some of the difference in performance between familiar and unfamiliar viewer
groups. Any experience that a familiar viewer had with any of these model’s appearance
changes would be have already been stored within the accepted range of appearances for
that model’s face. Thus, any variation, which a familiar viewer had been previously
exposed to for the faces in disguise, would not act as an effective disguise manipulation
for a familiar viewer, but it would for an unfamiliar viewer.

Although familiar viewers were somewhat impaired by evasion disguise, they could easily
see through imposter disguise. For an imposter disguise to be effective the model must
have moved both outwith their own face and into the face space of the person they are
trying to impersonate. Direction of disguise is limited to moving only towards (and into)
the face space of the other person, so disguise is automatically more difficult than in the
case of evasion whereby appearance can be changed in any direction that brings the face
image outwith its normally accepted face space. The familiar viewers were familiar with
both the imposter and target person who featured in each trial. This allowed viewers to
approach the task from two angles – the imposter face could be encoded to be the real
person behind the imposter, or it could be believed to be not a good enough match to the
target face to convincingly fall within their face space. Unfamiliar viewers were limited to
using only the second of these strategies. These factors help to explain why familiar
viewers perform with higher accuracy than unfamiliar viewers for both evasion and
impersonation disguise and also highlight reasons for evasion disguise causing familiar
viewers difficulty whereas impersonation disguise did not.

5.8 General Discussion
This chapter saw the creation of the FAÇADE database, which consists of images of male
and female human faces in disguise (evasion and impersonation) and no disguise image
conditions. Disguises were all model led, meaning that the models were free to disguise
themselves as they wished and disguises were not limited to the use of props. Image pairs

184

from the FAÇADE database were then used to test face-matching performance accuracy
for both familiar and unfamiliar viewers. In Experiments 10, 11 and 12 deliberate disguise
had an effect on overall face-matching performance accuracy – participants made more
identity judgment errors on disguise pairs than no disguise image pairs. Disguise pair type
also affected performance accuracy. Evasion disguises resulted in more errors than
impersonation disguise, and for unfamiliar viewers only (Experiments 10 & 11),
impersonation similar trials were more error prone than impersonation random trials.
These results show that disguise is more convincing when looking unlike yourself than
when looking like a specific other person. For the case of unfamiliar viewers,
impersonation disguises are more effective if the target looks similar to the model prior
to disguise. In Experiment 2, I found that unfamiliar viewers were no better at matching
disguised faces when they had been informed before completing the face-matching task
that some of the images would involve disguise, than they were when they were naïve to
the disguise element of the study (Experiment 10). In Experiment 12, a familiarity
advantage was found. Familiar viewers were better at correctly identifying match or
mismatch disguise face pairs than unfamiliar viewers (Experiments 10 & 11). Familiar
viewers easily saw through impersonation disguises, but even for familiar viewers,
evasion disguise caused significant impairment to face-matching performance.

Face matching performance for disguised face images was much lower than performance
from the GFMT, which is known as the standardised test of face-matching performance.
Viewers performed with around 89% accuracy on the long version of the GFMT and 83%
on the short version whereas here mean performance for all disguised faces was 77% for
unfamiliar (uninformed) and 80% for unfamiliar informed viewers. Furthermore,
familiarity has led to ceiling levels of performance in past face-matching tasks (e.g. Burton
et al. 1999), however performance of familiar viewers for evasion disguise was even
lower than that of unfamiliar viewers for cooperative stimuli in the GFMT. These
numerical comparisons highlight how severely disguise impairs face-matching
performance, with evasion disguise impairing even familiar viewers.

185

This study overcame the main limitations of past disguise research through use and
creation of the FAÇADE database and also the experimental design. For example,
Patterson and Baddeley (1997) Experiment 1, tested face memory for disguised images
against exact image memory performance. I test performance based on a disguise versus
no disguise face-matching accuracy comparison. The images included in the no disguise
condition were not the same image of a face. Additionally, the images used in Patterson
& Baddeley’s Experiment 1 disguise condition could have shown two different disguises
for the same identity – one at learning and a different disguise at test. I always compare a
disguise to no disguise image matching comparison rather than introducing two different
disguises. I believe having only one of the comparison images in disguise is better suited
to real world application.

The overarching result of this chapter is that disguise presence impaired face-matching
performance accuracy. This finding is in line with all previous research on disguise that
was discussed in the introduction of this chapter. Disguise had previously been shown to
impair face memory performance (Patterson & Baddeley, 1977; Terry, 1994; Righi et al.
2012). Dhamecha et al. (2014) also reported that face-matching accuracy for disguised
faces was lower than accuracy levels reported for the GFMT, which is a standard test of
cooperative face-matching. Dhamecha et al. (2014) tested matching for image pairs
whereby both of the images in the pair were generally disguised. My studies show an
effect of disguise on face-matching when just one of the images in the pair was in
disguise, showing how difficult disguise can make the task. Additionally, the stimuli used
by Dhamecha and colleagues (2014) included occlusion of features. Occlusion has been a
common feature of past disguise databases, and it seems obvious that face-matching
performance would be hampered when parts of the face are obscured. Testing facematching performance using the FAÇADE database showed that that even when disguises
do not occlude facial features, disguise makes matching faces a more difficult task than
when the same identities are presented undisguised.

186

This chapter advances upon previous disguise research through the investigation of
impersonation as well as evasion disguise. All past research discussed has looked at only
the case of evasion disguise (Patterson & Baddeley, 1977; Terry, 1994; Righi et al. 2012;
Dhamecha et al. 2014). I show that evasion and impersonation disguise both cause
matching difficulties, but evasion and impersonation do not cause equal levels of
difficulty. I was able to break these results down further as my design incorporated two
types of impersonation – similar and random. Unfamiliar viewers were better at matching
random disguise pairings than pairings where the model looked naturally similar to the
target they impersonated. Familiar viewers could see through both types of
impersonation. These findings highlight the importance of investigating impersonation
disguise and evasion disguise. The different results for each confirm that disguise cannot
be treated as one.

The findings from this chapter are also in agreement with previous face variability studies.
Past research has shown that integrating multiple images of the same identity (‘telling
people together’) is a difficult task (Jenkins et al. 2011). I have demonstrated that people
are worse at matching evasion face images (telling people together) than impersonation
faces (telling people together). As people are poor at integrating multiple images of a face
when there is not a deliberate attempt to make the identity look different across
photographs, it is not surprising then that performance is even poorer for same person
trials when there is a deliberate attempt to evade identity.

A difference however between the previous findings on face variability and the findings of
this chapter, relates to the effect of familiarity. Jenkins and colleagues’ (2011) familiar
viewers could easily group together multiple images of the same identity. In my facematching task familiarity did not completely overcome the difficulty of matching
disguised faces. Participants were significantly worse at matching face pairs where the
model evaded their own identity than they were at matching faces where there was no
disguise. Levels of performance for evasion disguise face pairs were also far lower than in
previous studies that have tested cooperative face-matching performance of familiar

187

viewers (Burton et al. 1999). As familiarity did not eliminate errors for evasion in my facematching task, this finding suggests that disguise took the faces outwith the area of
expertise that familiar viewers held for the faces involved.

It would have been interesting to test for graded effects of familiarity in this Chapter,
much like in Chapter 1. However, due to the participant samples that we used, familiarity
was skewed towards the extreme ends of our familiarity spectrums for each of our
experiments. Over 80% of the unfamiliar viewers rated the face items as completely
unfamiliar, whereas over 80% of our familiar viewers were of very or extreme familiarity
with the faces in the task. In addition to this, the scales were likely not directly
comparable as familiar viewers were more conservative on their familiarity ratings than
unfamiliar viewers due to their baseline level of familiarity with the faces. All familiar
viewers were at least somewhat familiar with all of the images as the models were of
their colleagues. This meant that those faces that qualified as extremely familiar were
usually who the viewers had an exceptionally high level of interaction with (e.g. in the
same office with, close friends with) where as people ranked lower down on the scale
would also be very well known, and likely classified as extremely familiar in many
scenarios, but were less familiar in relation to the other faces in the task. It is therefore
possible that even if a graded analysis was attempted on the small numbers in the lower
familiarity bands, that the graded result may not be found in this experiment given that
all familiar viewers were of a baseline high familiarity with the faces concerned.

This thesis has focused exclusively on conducting studies of face-matching performance
for various challenging situations. Face-matching is a current and common security
scenario, thus warranting thorough investigation, however it would also be interesting to
assess the situation of face memory for disguised and undisguised faces. Past studies
attempted to address this question but used image matching comparison scenarios and
questionable disguise stimuli. It would be interesting to find out whether face memory is
worse for disguised faces in our database. The case of impersonation disguise would be
particularly thought provoking as this has not received any past investigation and it is

188

unclear whether participants would be confused by seeing the target face or model in
undisguised form and be able to make links to the impersonator from the impersonation
disguise image or if they would incorrectly identify the target face as the face seen
before.

Future research could test human versus machine performance for the FAÇADE database
pairs. Previous research has tested machine performance for recognising disguised faces,
generally testing the performance of several different algorithms against each other (e.g.
Singh et al., 2009). Dhamecha and colleagues (2014) showed that machines performed at
a level similar to the unfamiliar human viewers for their face pairs that included faces in
disguise, but looked at only the case of evasion disguise and for an unrealistic set of
disguise faces. My database would allow a direct comparison between unfamiliar and
familiar human viewer face-matching performance with performance of computer
algorithms for evasion, impersonation similar and impersonation random faces. Human
performance has been successfully fused with algorithms in past studies (O’Toole et al.
2007) therefore if machine performance was tested on the FAÇADE database, fusing
performance from machines and humans may be a method which could be used to
improve disguise face recognition performance also.

Implications of this research are that care needs to be taken over disguise faces and
methods need to be explored to find out ways to improve disguise face recognition
performance. It would be interesting to test the performance of super-recognisers on the
task as super-recognisers have been found to outperform controls on other difficult tasks
of face-matching (Robertson, Noyes, Jenkins & Burton, 2016). Disguised face recognition
performance is poorer than performance for faces that are not disguised, hence more
errors are likely to be made in real life deliberate disguise fraud situations than if
someone is trying to pass themselves off to be someone else but has not made any
deliberate attempt to change their appearance to look more alike. Poor performance for
evasion disguise faces does however provide hope for undercover police investigations.
Police Scotland had addressed concerns about their undercover police officers’ identity

189

being recognised, but our findings suggest that evasion disguise is effective especially for
unfamiliar viewers. However the converse of this finding is that criminals who evade their
own identity provide a difficult face recognition scenario for the police.

I created a disguise face database, which is the first to include free evasion, and
impersonation disguises on real female and male faces. This has allowed, and will
continue to allow, disguise face recognition to be explored in more detail than ever
before. Disguise impaired face recognition performance, with the type of disguise
(evasion, impersonation similar or impersonation random) affecting face-matching
accuracy. Familiar viewers outperformed unfamiliar viewers at the disguise matching
task, however even familiar viewers were significantly affected by the presence of
evasion disguise. Implications of these findings are that familiar viewers will likely be
more accurate at tasks involving identity judgement than unfamiliar viewers. Poor
disguise face-matching is advantageous in terms of undercover policing but also poses
serious security threats if either evasion or impersonation disguises are used in criminal
situations.

Performance accuracy on the face-matching task has now been established, but it is not
yet known how people disguise themselves and which disguise manipulations work.
Disguised face images in the FAÇADE database were created from photographing models
posed in evasion, impersonation similar and impersonation random image conditions.
Importantly for these new research questions, models were free in their disguise. This
means that it was the models themselves who decided upon how to create each of the
disguises and props were provided and arranged on request of the models rather than
the experimenter. This free disguise element is unique to my database and allows an
investigation of what changes people naturally make to disguise themselves. Additionally,
to try and uncover the steps necessary to create an effective disguise, disguise facematching data from this chapter can be combined with written reports from both the
models and viewers regarding disguise manipulations and effectiveness. These questions
will be explored in Chapter 6.

190

Chapter 6 – Understanding Disguise
6.1 Chapter Summary
In order to further explore disguise - specifically how people disguised themselves,
whether effective disguises could be identified, and which disguises worked - data from
the matching task (Chapter 5) was analysed along with models’ records of how they
created their disguises and effectiveness ratings from an unfamiliar viewer group.

This chapter points to clear distinctions between evasion and impersonation disguise –
both in terms of what changes people make, and also what makes a disguise effective.
Unfamiliar viewers could accurately predict disguise effectiveness from viewing the target
and reference image side by side.

Previous studies on disguise investigated the effect of adding props to a face, which often
occluded facial features. I found that when models were free to create their own
disguises they used far more methods other than simply adding props to do this. Evasion
disguise revolved around creating differences with a target face in terms of internal and
external features, through the use of makeup, clothing and hairstyle change sometime
through wigs, and also using techniques such as expression and lighting change.
Successful evasion disguise was also linked to creating differences in social inferences.
Impersonation disguise on the other hand involved creating and focusing on similarities
with a target face, but these similarities were related to physical changes to a face, for
example internal and external features. Social inferences did not change to match those
of the person being impersonated. It is evident that free disguise creation, especially in
the case of effective disguise, is far more complex than purely the addition of props. It is
important that disguise research acknowledges distinctions between evasion and
impersonation disguise.

191

6.2 Introduction
As explained and demonstrated in Chapter 5, I have created a disguise face database
consisting of 26 models, disguised to evade their own identity and to impersonate the
identity of two reference individuals – one of naturally similar appearance to the model
and the other selected at random, from images of other models of the same gender.
These models were highly motivated, aided by the incentive of a performance based cash
reward, to create extremely convincing disguises. Models were free to disguise
themselves as they wished and could request props to assist their efforts. I provided no
guidance on how the models should disguise themselves, although I did alert them to the
different manipulations they could make using camera angle and distance (e.g. findings of
Chapter 4). The only limitations on disguise were that the end result must look like a
realistic I.D. photograph rather than a person in a fancy dress costume and that any props
that would have to be removed in a passport security check were disallowed (e.g. hats
and sunglasses). As demonstrated in Chapter 5 this led to the creation of a much more
sophisticated disguise database than previous studies have used. It is of interest to better
understand disguise - what the models did to disguise themselves in each condition,
which disguises were believed to be the best and what actually makes for the best
disguise (with reference to items that caused most difficulty in the matching task). I have
matching performance accuracy rates for the disguise face database (Chapter 5). I also
have information on what the models did to create their disguise and ratings from an
independent viewer group for the effectiveness of each disguise. This data can be used to
help to gain the desired better understanding of disguise.

As reviewed in Chapter 5, past research on disguise is limited in that it has not looked at
impersonation disguise, and models whose images provided the disguise stimuli used
simple disguises, adding props to occlude features. Perhaps the closest to background
literature as a starting point to predicting what may make the best disguise, is to look at
what image changes are already known to affect face recognition performance. Previous
memory studies focus on the case of same person face recognition, therefore apply more

192

to evasion disguise rather than impersonation. Most previous studies have focused on the
effect of changing expression, viewpoint or lighting on face recognition performance.

Factors that Increase the Difficulty of Face Recognition
The literature paints a clear picture that face recognition performance accuracy is worse
when viewpoint or expression is changed between presentation and test. Bruce (1982)
found that participants performed with 90% accuracy for correctly identifying an identical
picture at test as one that they had viewed 15 minutes earlier. If at test the participants
saw an image of the same identity that they had viewed before, but head angle or pose
was changed at the test phase, recognition accuracy dropped to 76%. Changes to head
angle and expression, rather than changes to just one or the other, led to further
impairments in recognition performance, with performance falling to 61% accuracy. The
effect of change in image between exposure and test has been shown to be so strong that
even familiar face recognition slows when viewpoint is changed (Bruce, 1982).

O’Toole, Edelman & Bulthoff (1998) report a similar effect. They demonstrated that
participants were less likely to recognise an image they viewed at test as an identity
previously seen if the conditions of the image view (full, 3/4 or profile) had changed
between the learning and test phase. When faces were learnt in full view, recognition
performance was highest for test items shown in full view. However, when faces were
presented in full view during the learning stage, there was an advantage for recognition
from 3/4 views over profile at test. This suggests that certain changes in viewpoint might
be more detrimental to recognition performance than others.

The studies mentioned above highlight that changes in viewpoint and expression
influence face recognition performance. Other studies have shown that these image
manipulations also affect unfamiliar face-matching accuracy (Bruce, Henderson,
Greenwood et al. 1999). Bruce and colleagues (1999) found that face-matching

193

performance was negatively affected by any change between the target and comparison
face images - this could include a change in viewpoint or change in expression. The
procedure took the form of a line up scenario. Participants compared a still image, taken
from video footage, to 10 face images in a line up array format shown below the target
video image. The task was to determine whether the target image was present in the line
up array, and if so, identify the correct match. Images in the line up array were always full
face and generally neutral in expression. The still video comparison images showed either
the same or different expression type (neutral or smiling) and were of either the same or
different view point (full or 30 degree view) to the target image. Accuracy was worse for
the trials where viewpoint or expression differed between target and comparison array
images. Viewpoint change was found to reduce performance accuracy more than a
change in expression.

Bruce et al. (1999) found that viewpoint could also influence performance on a paired
matching task. Participants made more errors when comparing a target face to a similar
distractor face if the viewpoint of the two images was different, than in situations where
viewpoint of the target and distractor were the same. Hill and Bruce (1996) mention that
when faces are matched across different viewpoints, performance accuracy is increased if
the faces are lit from the same direction, specifically if the images are both lit from above.
This is consistent with work of Johnston et al. (1992) which reported that lighting from
below made it harder to recognise familiar faces. Hill and Bruce (1996) reported that
changes in lighting alone reduced matching performance; as did changes in view point.
Together, these findings can be taken as evidence that a change in viewpoint, and also
expression or lighting, between target and comparison faces increases the difficulty of
unfamiliar face-matching. Any consistencies held across the images can aid the matching
effort (e.g. Hill & Bruce, 1996).

194

Changes to Internal Versus Changes to External Features
A second question of interest on the topic of which changes matter, is whether people
pay different amounts of attention to internal and external features when making identity
judgments. If they do, this may give us some clues to predict whether internal or external
changes matter more for effective disguise. The answer appears to depend on whether
the viewer is familiar or unfamiliar. Unfamiliar viewers value internal and external
features as equally important when matching a face (Ellis et al., 1979; Young, et al., 1985).
Familiar viewers on the other hand give more weight to internal features. Notably these
are the less changeable features of a face (Ellis et al., 1979; Young et al. 1985; Tanaka &
Farah, 1993; Toseeb, Keeble & Bryant, 2012). While hairstyle can change frequently,
specific features of a face are generally more constant. Young et al. (1985) found these
results by testing how quickly participants made correct matches for a full face image
alongside only the internal features, compared to matching a full face image with an
image cropped to contain only external features. Features themselves (internal and
external) may be easier to compare across images taken from the same viewpoint (Bruce
et al. 1999).

O’Donnell & Bruce (2001) report a slightly different but interesting finding for familiarity
and attention to internal and external features. They artificially edited images of faces so
that participants would see either two identical images, or the original image and an
altered image of that same face side by side. Changes could be either to the hair or chin
(external features) or the eyes or mouth (internal features). Participants had to work out
whether the images in the pair were identical or differed physically in some way.
Unfamiliar viewers (untrained) detected changes to the hair with highest accuracy (in line
with findings of Bruce et al. 1999). Familiar viewers (trained) were very good at detecting
changes to the hair, but were also highly attuned to detecting changes made to the eyes.
This experiment made artificial changes to images resulting in an image matching type
task rather than face-matching using disguise manipulations. In conclusion, familiar
viewers are consistently relying on internal features of a face to aid their identity
decisions. It is uncertain whether unfamiliar viewers are using internal and external

195

features equally (Ellis et al., 1979, Young et al. 1985), or relying more on external features
(Bruce et al. 1999; 2001). Evidence can be found for both accounts in the literature.

Past Research on What Makes an Effective Disguise
Patterson & Baddeley (1977) begin to answer the exact question of interest, of which
disguises work. As is the case for all previous disguise research, this situation was
investigated for evasion disguise only. Unlike the findings above, they reported that
recognition performance was not greatly affected by small changes in appearance. They
specifically examined changes to pose and expression between learning and test items.
Although participants showed a lower overall hit rate when the test items were of
changed view and expression than identical view and expression, their false positive rate
was also lower, negating any overall effect of changing viewpoint and expression. The
study also tested recognition accuracy for disguised faces. Many recognition errors were
made for the disguise stimuli – hit dropped rate from .98 for no change photographs to a
hit rate of .45 (around chance level) for the disguise images. Disguise images in this
experiment in Patterson & Baddeley’s study were taken from actor photographs that
showed actors in a range of different appearances depending on the roles they had taken
on. Evasion or impersonation disguise itself was not directly the aim of the change in
appearance. Actors had changed their appearance to suit a job role. I return to the issue
of changing character later in this chapter. Active disguise manipulations, whereby
someone is specifically trying to evade or impersonate identity, may be more challenging
for face recognition.

Due to the type of disguise stimuli used in Patterson and Baddeley’s Experiment 1, they
were unable to categorize the exact changes made to appearance across images. Like me,
Patterson & Baddeley wanted to understand better what appearance changes made for
an effective disguise. This was addressed in Experiment 2 of the paper. In Experiment 2
stimuli were amateur dramatic students who modelled for disguise and no disguise
photographs. However, unlike in my FAÇADE database, their disguises were standardised,

196

and were limited to the addition or removal of wigs, beards and glasses. All models were
male. Participants had been informed that the appearance of the models might have
been changed (through disguise) in the test images, although in Experiment 11, I saw that
knowing to look for disguise did not improve performance compared to those who were
not informed of the disguise manipulation. The effect of each of disguise was analysed,
with the findings as follows: matching performance was greater when hairstyle remained
the same than when it was changed; performance was better when a face was unchanged
with regard to presence or absence of a beard; performance was poorest when multiple
changes to appearance were made. The presence/absence of glasses interacted with
changes in hair. The general finding was that disguise made it harder to recognise faces,
and multiple disguise manipulations (more props) increased the difficulty. It is a recurrent
theme that multiple alterations to a face between learning and test images have greater
effect on performance accuracy than just one change, or fewer changes (Patterson &
Baddeley 1977; Bruce et al. 1982; Bruce et al. 1999).

Dhamecha et al. (2014) were interested in specific elements of disguise that cause
problems for face-matching performance. As explained in the introduction of Chapter 5,
Dhamecha and colleagues looked at the effect of occluding facial features through the
addition of props and then cropped the images so that the focus was exclusively on
internal features of a face. Following collection of results from their matching task,
Dhamecha and colleagues (2014) analysed performance accuracy for each segment or
segments of the face covered. This was achieved by dividing the face into four regions:
lips, nose, eyes and forehead. The occlusion (presence of a prop) of just one of any of
these four regions was linked to high error rates in the face-matching task. However it is
important to remember that due to the design type, a one-region disguise pair could
consist of one face where the eyes are not visible due to sunglasses next to a face where
the lips were not visible as they were covered by a medical mask (see Figure 6.1). This
would leave very little of the face available for comparison across images, so it is little
wonder that this is a highly error prone task. The presence of multiple props (covered
areas) led to some further reduction in performance, but the greatest reduction in

197

performance came from the introduction of any single occlusion (i.e. from zero to one
occluding disguise element).

Figure 6.1 Figure demonstrates two different disguise images for the same identity. The image on the left
occludes top part of the face, and the image of the right occludes the bottom part of the face.

All of this previous work has covered only situations of evasion disguise – manipulations
that make a person look less like themselves. The situation of impersonation disguise -
taking on the identity of a specific other person - has not been explored previously, thus I
will be examining what makes for an effective impersonation disguise for the first time.
Previous work has reported that distinguishing features are important for face
recognition, with identities in a line up task being more likely to be mistaken for the true
culprit if they share a distinguishing feature (Wells, et al., 1993; Flow & Ebbesen, 2007).
Therefore this might provide a clue for creating convincing impersonation. Other than
this, all that is predicted for impersonation disguise is that effective impersonation
disguises will be those that make a face look most similar to the target face.

In summary, so far it has been established that any change between images presented at
learning and test causes difficulties for recognition performance, with disguise
manipulations (here seen as adding or removing props such as wigs, beards or glasses)
causing more difficulty than more subtle changes such as expression and pose (Patterson
& Baddeley, 1977; Bruce et al, 1992, 1999; O’Toole et al. 1998). Expression and pose

198

differences between images have been found to impair matching performance (Bruce,
1999), suggesting that these changes could assist the success of disguise. Disguise, in
terms of occluding facial features (Dhamecha, 2014), and also free disguise (Chapter 5),
leads to severe difficulties in face-matching.

In Chapter 5, I applied theories to both evasion and impersonation disguise. For an
evasion disguise to be successful an individual must change their appearance so that it
becomes outwith the accepted range of appearances for that face. In the case of
impersonation disguise the challenge is to change appearance to get inside a specific
other person’s accepted range of appearances. It thus seems likely that the approach
taken to create each of these types of disguise (evasion and impersonation) may be
different. It is less clear whether differences will emerge between the approaches to
creating the different types of impersonation disguises – impersonating someone of
similar appearance compared to impersonating an identity of the same selected at
random, but of the same sex. It is possible that natural similarities and differences
between specific faces may guide each individual’s approach, leading to a very individual
method for disguise, or alternatively there may be a common approach taken by our
disguise models to create impersonation disguises, for example models might consistently
focus on hair or facial expression. I will approach two key questions – how people disguise
themselves and which approaches are effective.

6.3 How do People Disguise Themselves?
Method
Participants
Three independent raters were instructed of the coding system and any coding
discrepancies discussed until agreement was met.

199

Design
Word clouds were created for the three different disguise conditions – Evasion,
Impersonation Similar and Impersonation Random.

Procedure
The 26 models recorded the changes that they made to their appearance by writing down
on a piece of paper what they did to disguise themselves in each condition and what they
hoped each manipulation would achieve. This was recorded whilst the models created
their disguises.

The disguise manipulations and purposes were typed up and coded into categories to
make sure the same changes, described by different words, were captured under the
same word to allow a word analysis to be accurately conducted. For example, if a
participant dyed their hair, this was categorised as a change of hair-colour. Categorisation
was conducted by 3 raters who had been instructed of the coding system. Any
discrepancies were discussed amongst the raters until agreement was met.

Hyphenations were used to make sure that meaningful phrases, such as hair colour, were
kept as one word. Otherwise the word hair for example, would have been counted across
several different hair changes - such as hair colour and hair length. Hyphenating words
that contributed to a phrase ensured accurate word categorisation.

The categorised body of text was entered into wordle (http://www.wordle.net) –
software that creates word clouds based on word frequency. Words are given
prominence within their cloud based on the frequency of their occurrence. The more
often a word occurs in the inputted text, the larger the word will be displayed in the word

200

cloud. These word clouds may help in gaining knowledge of the manipulations made by
the models for each of the disguise conditions.

Results and Discussion
Evasion

Figure 6.2 Word cloud showing the most frequently stated words for creating an Evasion disguise.

All changes in the Evasion condition related to creating differences in appearance
compared to the model’s own reference photograph. Models frequently changed the
appearance of their own features, skin-tone, hairstyle or hair colour and clothes in order
to look make themselves look physically different to their normal (own reference
photograph) appearance (see Figure 6.2). The word cloud includes ways that participants
tried to change their identity and shows that this was largely done through props, such as
use of the words ‘wig’, ‘glasses’ and ‘clothes’. Props were however not the only method
used to create facial change. Other techniques used included the use of makeup to
facilitate changes to features and skin tone and changes in camera angle to change face
shape.

201

Impersonation Similar

Figure 6.3 Word cloud showing the most frequently stated words for creating an Impersonation Similar
disguise. All words represent similarities with the target face except where specified as differences.

To create Impersonation Similar disguises models focused on looking similar to, or even
matching the appearance of, the reference photograph. This was attempted in a variety
of ways including focusing on creating similarities in the appearance of specific features
especially hairstyle but also eyes and lips (see Figure 6.3). Beards were grown or shaved
where applicable to match the condition of the reference photo (see Figure 6.4). Copying
the eyebrows of a face was a particularly common impersonation similar disguise
technique (see Figure 6.5). This was done by changing the eyebrow colour, covering up
eyebrows with foundation, or drawing on eyebrows in a different style or shape to the
model’s own to create a better eyebrow match with the reference photo. Copying the
facial expression of that in the reference-photo was also a very popular impersonation
technique (as in both Figure 6.4 & 6.5).

202

Figure 6.4 The model (shown right in this image pair) shaved his beard to better match the appearance of
his target (left).

Figure 6.5 The model (right) has copied the eyebrows of the target (left) using makeup to alter eyebrow
shape.

203

Impersonation Random

Figure 6.6 Word cloud showing the most frequently stated words for creating an Impersonation Random
disguise. All words represent similarities with the target face except where specified as differences.

For the Impersonation Random condition, models took a roughly similar approach to their
disguise as they did in the Impersonation Similar condition. Models again focused on
creating similarities with the reference-photo and matching on a range of features. It is
interesting to note that hair, an external feature, became the predominant focus of
change, and there was slightly less focus on internal features (see Figure 6.6). The
changes to internal and external features were somewhat more balanced in the case of
impersonation similar disguises. Perhaps models felt it harder to manipulate internal
features when there were less natural similarities, additionally there could be more
differences between hair style in the random matching to account for. There was also a
little less focus on copying facial expression than there was in the impersonation similar
condition. This may be due to expression matching resulting in a picture matching type
effect, whereby the impersonation random models could be revealed as an impersonator
if viewers spot slight differences in the copied pose.

The most common approach to impersonation disguise fits with the theoretical
framework adopted in this thesis. Impersonation disguises rely on successfully creating a
face image that would be accepted as falling within the other person’s face space. The

204

models make a clear effort to do this, as they focus on creating similarities with the
reference photograph. It is unknown at this stage whether the changes that were most
common were actually those that caused most difficulties for matching. This will be
looked at in section 6.5. First I will discuss whether viewers can tell by eye which disguises
will be effective.

6.4 Can Viewers Predict by Eye which Disguises will be Effective?
The next question of interest is whether the images rated to be the best disguises were
those that caused most difficulties in the matching task. This investigation will help to
understand if viewers can determine disguise effectiveness simply through a side by side
comparison of the target and disguise face images (evasion, impersonation similar,
impersonation random).

Method
Participants
The independent viewer group (first mentioned in Chapter 5), who chose the final stimuli
for the FAÇADE database, also provided the participant group for this study.

Procedure
The independent viewer group provided disguise effectiveness ratings for each of the
selected database disguise images on a scale from 1-7 with 1 being a very poor disguise
and 7 an extremely convincing disguise.

205

Design
To find out whether the viewer group could accurately predict disguise effectiveness, I
compared the effectiveness ratings for each of the disguise images as scored by the group
of four unfamiliar viewers, to performance accuracy for each of the items on the facematching task. This was conducted for each of the three disguise conditions – Evasion,
Impersonation Similar and Impersonation Random.

Results
Pearson correlation coefficients between ratings of by disguise effectiveness of the face
images and performance accuracy for the disguised images on the face matching were
calculated to find out whether effectiveness ratings predicted matching accuracy.

Figure 6.7 Graph showing correlation between effectiveness rating and percentage of errors made for each
Evasion disguise item. Data points are spread horizontally if they would otherwise overlap.

206

Figure 6.8 Graph showing correlation between effectiveness rating and percentage of errors made for each
Impersonation Similar disguise Item. Data points are spread horizontally if they would otherwise overlap.

Figure 6.9 Graph showing correlation between effectiveness rating and percentage of errors made for each
Impersonation Random disguise item. Data points are spread horizontally if they would otherwise overlap.

207

Disguise ratings were strongly correlated with performance on the face-matching task for
both Evasion disguises r =.66, p<.01 (see Figure 6.7) and Impersonation Random disguise r
= .65, p<.01 (see Figure 6.9). For these conditions the higher the effectiveness rating of
the disguise, the more face-matching errors were made. Correlation levels were almost
significant in the disguise impersonation similar condition, r = .39, p=.05 (see Figure 6.8).
The Impersonation Similar condition is naturally a more difficult condition for the raters
to judge based on the similarity of the target and reference photograph. There was
however a highly significant correlation for impersonation when collapsing across
impersonation type, r = .54, p<.001.

These results suggest that the effectiveness of a disguise can be judged quite accurately
by simply showing a group of viewers both the reference photograph and disguised image
side by side and asking them to rate how good the disguise is.

6.5 What do Viewers Believe Makes for an Effective Disguise?
Now I have shown that viewers can accurately predict which disguises would be effective,
I believe it is interesting to explore what it was about a disguise which the viewers
thought made it effective. It is believed that as the viewers were accurate at making their
disguise effectiveness decisions, their insight into what made an effective disguise will
also be useful. Obtaining an understanding of what makes an effective disguise may make
it easier to create successful disguises in the future; this could have useful applied value
such as aiding the disguise of undercover police officers.

208

Method
Participants
The participants were again the 4 raters who made the stimuli selection decisions
discussed in Chapter 5, and also gave the disguise effectiveness estimates used in the
study above.

Design
Effective disguise changes were categorised and counted for each of the three disguise
conditions.

Procedure
As part of the stimuli selection phase, four unfamiliar viewers worked collaboratively to
decide upon the best match or mismatch image from a range of options for each model
for each disguise condition. During this process I asked the viewers to rate how good they
thought that each of the disguises were for each of the selected images on a scale from 17. A score of 1 indicated that the disguise was very poor, whereas a score of 7 meant that
the disguise was extremely effective. These were the ratings used for the correlations in
section the study above. In addition to these ratings, the viewers provided comments on
what they thought made the chosen disguise images effective in each of the conditions.

The comments made by the viewers were coded into seven categories of change –
internal features, external features, expression, skin-tone, social inferences, face shape
and other.

209

Results & Discussion

Figure 6.10 Bar graph showing the most frequent forms of disguise for Evasion (blue), Impersonation
Similar (red) and Impersonation Random (grey). Evasion changes capture differences in appearance with
the reference photograph whereas Impersonation changes represent similarities.

Figure 6.10 shows the frequency of effective changes by category for each of the three
disguise types. From this figure it is evident that effective Evasion disguise was created by
many different types of changes. Changes to Internal (specific features of the face) and
External changes (changes to the hair and also any addition of props, including clothing)
were made to the face (specifically changes to nose shape, eyes and hair), but changes in
Expression, Skin-tone, Personality, Face-shape and Other changes (e.g. changes in
lighting) also contributed greatly to effective Evasion disguise. Whereas the models
tended to focus mostly on changing hair and makeup when creating their disguise (Figure
6.2), interestingly the viewer group also picked up on the personality differences and
changes to face-shape that these changes evoked and found these to be effective factors

210

of change to create evasion disguise. For example, viewers described a disguise as
effective if they thought that the images represented different characters, such as one
image showing a quiet and studious person and another showing a more outgoing and
party loving person (see Figure 6.11).

Figure 6.11 Image example where social inferences were reported to differ between the reference model
image (left) and the model in evasion disguise (right).

Using the numerical information entered into the bar chart above, a binomial test was
conducted to find out whether there was a significant difference in the proportion of
Feature to Non-Feature changes for Evasion disguise compared to Impersonation
disguise. The test proportion was calculated from the proportion of Feature (represented
as internal features and external on the graph) and Non-Feature (all other categories of
change from the graph) changes made for Evasion disguise, and this was compared with
the same proportions for Impersonation disguises. The binomial test indicated that for
Impersonation Similar disguises the proportion of .69 feature responses was higher than
the expected .34, p <.001. This same pattern of results was found for Impersonation
Random disguises, with the proportion of .67 features being higher than the expected
.34, p<.001.

These binomial results highlight that whereas effective Evasion disguise encompassed
many different changes, across many different categories, with other changes

211

outweighing internal and external changes, effective Impersonation Similar and
Impersonation Random disguises were dominated by the internal and external feature
changes. The bar graph shows that internal featural similarities were the most common
effective change for impersonation similar disguises (similarities between the nose, ears,
eyes, forehead, eyebrows & mouth). Changes to external features followed, these were
generally related to hair. Personality and expression were much less noted for creating an
effective impersonation than they were for evasion. These results suggest that to create
an effective impersonation similar disguise the focus should be on creating similarities
between the internal features of a face. The effective Impersonation Similar
manipulations were very similar to the effective Impersonation Random manipulations.
Internal features again came out as most important, followed by hairstyle changes
(External). For the impersonation conditions, models sometimes tried to copy any
distinguishing feature of the person they were trying to impersonate. This made the
disguise more effective and was captured under the category ‘Other’ (see Figure 6.12 for
example). The presence of distinguishing features has been found to affect face matching
accuracy in past studies of facial recognition (Wells, et al.,1998).

Figure 6.12 The model (left) copied the distinguishing feature (mole [on the left side of the image under the
mouth]) of the target (right) by using make up.

212

More changes, and categories of change contributed towards effective Evasion disguise
than Impersonation disguise. Effective Impersonation disguises generally revolved around
changes to internal and external features of a face whereas Evasion disguise also included
changes in perceived face-shape and personality, as well as ‘other’ manipulations, which
were those not captured by the specifically labelled categories. It is important to
remember that previous research has been limited to looking at effective Evasion disguise
only, and even within that Evasion disguise was investigated in relation to the effect of
specific props and occlusions of certain areas of the face. No features were occluded in
my disguise database. Previous studies have missed most of the action by focusing only
on the category that I named External. My results show that far more goes on in the
creation of free disguise, for both Evasion and Impersonation, than simply the addition of
props, such as wigs and glasses.

6.6 Experiment 13 - Do Social Inferences Change for Disguise?
The graph and binomial analysis above demonstrated that the viewers’ disguise
effectiveness verdicts were influenced not only by the featural types of change explored
in previous research, but also by aspects relating to social inferences. This was especially
true in the case of evasion disguise. I believe this is an interesting topic to investigate
further – most research on disguise has used props to change appearance. Never before
in the literature has disguise been studied with the intention of investigating disguise
related changes in perceived personality traits.

Patterson & Baddeley (1977) showed that encoding faces in terms of social inferences
was a powerful tool to aid facial recognition. In their experiment participants viewed face
images during a learning phase and were instructed to make judgments on either the
features or personality of the person in the image. Participants who were assigned to the
feature condition made ratings of nose size (small – large), lip size (thin – full), width
between eyes (close together – far apart), face shape (round – long). Those instructed to
make personality judgments rated niceness (nice – nasty), reliability (reliable –

213

unreliable), intelligence (intelligent – dull) and liveliness (lively – stolid) of the individuals
shown in the images. In both of the conditions participants were informed that making
these judgments of the face might help them to remember the faces. It was found that
those participants who made personality trait inferences for the faces were better (but
not significantly so) at recognising faces at test than those who made feature related
judgments. This trend provides an interesting hint that personality perception may
influence the encoding of faces, but cannot be taken as evidence as a significant result
was not found. The sample size was fairly small at 18 participants in each condition; a
larger sample size may have provided a significant result. Combined with the comments
received from my viewer group regarding what makes a good disguise, it is possible that
changes in social inferences may have influenced performance on the face-matching task
of Chapter 5.

Trait perception research has generally considered whether accurate social inferences
about a person can or cannot be acquired from viewing a photograph of that person’s
face. Those studies have typically considered natural images, meaning that the
photographed person has not been asked to express any particular trait, but instead may
or may not reflect aspects of their true personality unintentionally in photograph. To find
out whether viewers’ judgments of personality are accurate, scores from personality
questionnaires taken by the models are generally compared with viewers’ trait scorings of
the photograph. Several studies have found that above chance level judgments can be
made in this way. For example extraversion was accurately judged from face images
viewed for as little time as 50ms (Borkenau, Brecke, Möttig & Paelecke, 2009). This social
judgment was linked to the display of joyful facial expressions such as smiling (Kenny,
Horner, Kashy & Chu, 1992; Borkeneau et al. 2009). Personality judgments recorded by
Naumann, Vazire, Rentfrow & Gosling (2009) were accurate in both an uninstructed
natural photograph and a neutral pose photograph condition, with the authors reporting
that expression was not the only factor to aid social inferences from a photograph.
Factors such as clothing were reported to also lead decisions, as well as general
appearance and the way people held themselves. Rule & Ambady, 2011 even suggest that
earlier life photographs can suggest success in later life, with power inferences from

214

college yearbooks predicting later leadership success. These findings portray social
inferences as a stable concept attached to an individual, which holds across photograph
and time. If this is the case, then social inferences could actually help in identity judgment
tasks.

Whereas most past research looked at the ability to take accurate social inferences from
photographs and perceive these inferences as a stable judge of character, in some
situations people may want others to infer inaccurate social inferences from their
photographs. For example appearing to be a better candidate for a job or enhancing
likeability on a social media account or online dating profile. Leikas, Verkas & Lonnqvist
(2012) found that it was possible for the same person to have different social inferences
made about them across different photographs. Models were asked to change in
appearance to match high and low ends of the spectrum on each of the Big 5 Personality
Traits, and photographed in each of these conditions – therefore the study focused on
deliberate changes in appearance rather then incidental change between different
photographs of the same face. Each of the Big 5 Personality traits was described by two
adjectives that the models could use as a guide for creating their appearance in the
corresponding photograph. For example ‘anxious and stressed’ described the trait
neurotic, whereas in the stable condition models were asked to make themselves look
‘stable and quiet’. Models were limited in the ways that they could achieve these
personas, with the paper stating that ‘targets were not allowed to add, change or remove
clothing, hairbands or decorative items, to remove or add makeup, or to groom their hair
between conditions’ (Leikas et al. 2012). Extroversion could be convincingly changed
across photographs, and to a lesser degree neuroticism and conscientious. Agreeability
was not successfully changed across posed photographs. These findings suggest that it is
possible to deliberately control social inferences in a photograph. I may see even greater
changes in the disguise experiment as my disguise models were given far greater freedom
regarding what they could do to change appearance, and as shown in the section above,
the models used a great range of strategies to change their appearance in addition to
purely pose and expression.

215

It is possible, and seems viable based on the findings of section 6.5, that social inference
may change as a result of disguise manipulations. Some incidental change is expected
across multiple undisguised images of the same face. In this next experiment, I will test
whether social inference judgments differ more for disguise change than naturally as a
result of incidental change. The traits that will be explored are Trustworthiness,
Attractiveness and Dominance. These traits have been chosen as they are the 3 traits that
people spontaneously characterise faces on, according to PCA analysis (Todorov, Said,
Engell & Oosterhof, 2008; Sutherland et al. 2013).

For the case of evasion it is predicted that there will be a greater difference in social
inferences for disguise related change (different in social inference judgments between
the reference photograph and disguise photo) than for incidental change (difference in
social inferences between two undisguised images of the model). As models creating
evasion disguise are trying not to look like themselves in their evasion disguise image,
creating a situation where their social inferences are changed in relation to those made
for a no disguised reference image may help to hide their true identity. For impersonation
disguise models tried to look more similar to the target, this may have been reflected in
similar social inferences. For the case of impersonation disguise it is predicted that
differences in social traits will be larger for incidental change (the target photograph and
undisguised model image) than between the impersonation disguise image and the image
of the target person).

Methods
Participants
30 undergraduate students (M = 12, mean age = 20.4) from the University of York (who
had not taken part in any previous experiments involving the FAÇADE data set)

216

volunteered as participants in this experiment in exchange for a small cash reward or
course credit.
Stimuli & Design
The stimuli were all the images from the FAÇADE image database, presented one at a
time. For each model the following images were shown: reference, no disguise, evasion
disguise, impersonation similar disguise, target similar person’s reference, impersonation
random disguise and the target random person’s disguise reference image. All images
were presented in a random order.

This is a within subjects design whereby each participant rates each face image on each of
the three Traits (Trustworthiness, Attractiveness and Dominance). Each of the Traits were
rated in a random order for each face image.

Procedure
Participants viewed the stimuli (one image at a time) on a computer screen. Underneath
each image appeared the question, either ‘How trustworthy is this person’, ‘How
dominant is this person?’ or ‘How attractive is this person?’ The image remained on
screen until all 3 of these questions had been answered for the face, just one question
was shown at a time and the questions were always asked in a random order. Participants
indicated their response to each question by using the computer mouse to select an
answer on a rating scale between 1 (low score on the trait) and 7 (high score on the trait).

Results and Discussion
Social inference distance change was always measured according to the square root
distance between face images rather than the distance between means. Square root
distance took into account the fact that some people naturally looked more similar than
others did to the person they were trying to impersonate on certain traits. The square

217

root distance was calculated for disguise to target face images, and compared to the
incidental change for the corresponding no disguise images.

Figure 6.13 Example illustration of the distance calculations made for the Evasion disguise condition.
Distance moved for incidental change was compared with distance moved for disguise change for each of
the 3 disguise conditions (Evasion, Impersonation Similar, Impersonation Random).

In cases of evasion people moved significantly further from their perceived personality
traits (reference image) compared to the change in trait perception for another no
disguise image of that person, t (25) = 7.71, p <.001, CI = 1.19 to 2.06. The mean distance
moved in trait perception between reference photo and no disguise image was 0.63, SD =
0.38, SE = 0.08. The mean distance moved in trait perception between reference photo
and evasion photo was 2.25, SD = 0.92, SE = 0.18.

To investigate where these differences in traits lay, results were analysed as above, but
this time for each social inference (Trustworthiness, Dominance and Attractiveness) in
turn. There was a significant difference in perceived trustworthiness scores between the
reference and no disguise self images M = .34, compared with the difference between the
reference and evasion images M = 1.43, t(25) = 6.54, SE =.16, CI = 0.74-1.43, p<.001.

218

There was also differences in perceived dominance scores between reference and no
disguise self images M = .24, compared with the difference in perceived personality traits
for the reference and evasion images M = .60, t(25) = 3.55, SE =.10, CI = 0.14-0.56. p<.005.
Finally, there was a significant difference in perceived attractiveness scores between the
reference and no disguise self images M = 0.36, compared with the difference in
perceived personality traits for the reference and evasion images M = 1.42, t(25) = 5.63,
SE =.19, CI = 0.67-1.44, p<.001. These results show that the there were significant
differences in social inferences for each inference in turn as well as an overall effect. This
was as expected based on the results of the effective disguise section earlier, where
personality change was listed as an effective form of disguise.

Impersonation Similar
For impersonation similar faces there was not a significant difference between perceived
social inferences as rated as a whole for the reference and impersonation images (Mean
distance = 1.19), compared with the difference in perceived personality traits for the
reference and no disguise self images (Mean distance = 1.24), t(25) = .38, CI = -0.31 –
0.21, p =>.05, SE = .13. This was the case for each inference individually and overall (see
Table 6.1).

Distance between

Distance between

target reference

target reference

image &

image & model’s

impersonation

reference image

Significance Level

image
All traits combined

1.19

1.24

p>.05

Attractiveness

.73

.82

p>.05

Dominance

.38

.42

p>.05

Trustworthiness

.67

.60

p>.05

Table 6.1 Social inference comparisons for impersonation similar images.

219

Impersonation Random
The pattern of results for impersonation random items was similar to that for
impersonation similar items – no results were significantly different at either an overall
level or trait breakdown (see Table 6.2).

Distance between

Distance between

target reference

target reference

image &

image & model’s

impersonation

reference image

Significance Level

image
All traits combined

1.31

1.24

p>.05

Attractiveness

.75

.97

p>.05

Dominance

.51

.55

p>.05

Trustworthiness

.75

.66

p>.05

Table 6.2. Impersonation random disguises, means and median results for each analysis.

The results for impersonation disguises were not as hypothesised – impersonation
disguises did not move a target face significantly closer to the trait perceptions of the
model. This is perhaps not entirely surprising, as similar personality was not picked out as
particularly important in the effective disguise section above. It seems that in the case of
impersonation, other factors, such as featural change may be more important than
changes of social inference.

Discussion
Social inferences differed significantly in the case of evasion disguise, suggesting that a
change in these perceptions occurs when people are trying to not look like themselves.
Social inference must be based on physical appearance. The physical appearance changes

220

made to create Evasion disguise influenced social inferences. Judge of character is
generally believed to be accurate, but people are fooled by social inference change in
terms of identity judgment in the case of evasion disguise.

Impersonation disguise on the other hand does not rely on the match of social inferences
between a model in impersonation disguise and an image of the target person that the
model was trying to impersonate. In this scenario other factors of similarity seem to be
more important e.g. specific and distinguishing feature match. Instead of trying to mimic
the target image itself, e.g. the exact pose and expression, models may instead be
attempting to portray another way that the target identity could appear, e.g. a different
pose and expression. It has already been established that it is harder to match identities
over changes in expression (Bruce et al. 1982), than in unchanged expression. This may
work to an impersonator’s advantage in the case of disguise, and also explain why social
inferences are not matched for impersonation similar disguise. If for example, a model
chooses to try and look like the target person when they are in a happier or angrier mood,
the social inferences drawn from the face may differ even if changing the expression
makes the images look more similar in identity overall.

6.7 General Discussion
In summary, the disguise models applied many different manipulations to create each of
their disguises. Viewers were able to predict, by looking at a target image and disguised
comparison image, whether the disguise was effective or not (with relation to the
percentage of errors made for that face in the face-matching task Experiment 10).
Furthermore, manipulations that make an effective evasion disguise are different to those
that make an effective impersonation disguise. Whilst all disguises involve either the
similarity or difference in internal and external features, evasion disguise also
encompassed more non-feature factors including changes in expression and perceived
personality. Finally, Experiment 13 confirmed that social inference change occurred only
in the case of evasion disguise.

221

Evasion disguises involved internal features and external changes, with internal changes
often being achieved through the use of makeup and also changes in expression.
Impersonation similar changes involved creating similarities with the target face which
included similarities in hairstyle, internal features, clothes and expression.

Impersonation random disguises focussed mostly on creating similarities in external
features. The most important message here, with relation to previous work, is that
models did far more to create their disguises than simply the addition of props. Prop
additions and occlusion of features with props have been the most common method of
creating disguise stimuli for disguise investigation in previous research. My research
shows that when given the freedom to create their own disguises, models use many more
disguise techniques other than simply the addition of props. If disguise research is to have
real world relevance, then free disguise manipulations should be allowed when
constructing model stimuli.

Additionally, this chapter reiterates the differences between Evasion and Impersonation
disguise. Chapter 5 highlighted that Evasion caused more matching difficulties than
Impersonation. This chapter goes further by showing that the approach taken to these
disguises, and the factors which make them effective, differ for evasion and
impersonation. Whereas both disguise types include featural manipulations made by the
models, the independent viewer group picked out more changes in personality and
expression as factors that made a disguise effective for Evasion disguise than in either
case of Impersonation. In line with this finding, social inference change occurred only for
Evasion disguise. These results further highlight the importance of a distinction between
evasion and impersonation when investigating disguise.

222

One interpretation of these findings could be that large differences in appearance tend to
make ‘same’ responses unlikely and be accompanied by changes in social inferences.
Because Evasion disguises are less constrained than Impersonation disguises, they can
result in major image differences. That leads to higher error rates (as seen in Chapter 5)
and distinct social attributions.

It seems that effective disguise does not revolve around a simple recipe that would work
for all faces. Lots of different changes are taking place, and some of these changes will
work better for some individuals than others depending on the natural appearance of a
face. What is evident is that there is more to disguise than simply the addition of props
and furthermore Evasion and Impersonation disguise are achieved in different ways.

223

Chapter 7 – General Discussion
7.1 Overview of Findings
The research reported in this thesis investigated face recognition in challenging situations.
The introductory chapter outlined that the critical task of face-matching is difficult for
unfamiliar viewers, yet trivial for familiar viewers. These findings came from past studies
that used cooperative stimuli, meaning that for same face pairs, the person who was
photographed made no deliberate attempt to change their own appearance across
multiple images, and images were taken under constant conditions with extremely short
time intervals between photographs. In cooperative stimuli tasks, different face trials
paired the most similar faces from a small pool of available images (Burton et al, 1999;
Bruce et al. 1999). In this thesis I argued that performance is likely even worse for images
of a challenging nature – images that include incidental or deliberate face variation or
across identity similarities and reduced image quality. I explain that there may even be
limits to the familiarity advantage (times when familiarity can not completely compensate
for poor performance). I investigated face-matching situations where the same face
image pairs incidentally looked different due to within person variation across ambient
images, and also where different identity pairs were of extremely similar appearance due
to natural facial similarities between celebrity and lookalike faces (Chapter 2).
Investigation continued for unintentional appearance change due to change of camerato-subject distance (Chapter 4). And finally I examined matching performance for
deliberate disguise, where an individual deliberately made changes to evade their own
identity or to impersonate someone else (Chapter 5). The manipulations made to create
the disguise face-matching stimuli were explored (Chapter 6). I also tested ways of
improving performance for challenging face images (Chapter 3). In total, 16 studies were
conducted with the aim of furthering understanding of face recognition in challenging
situations.

The investigation began by testing face-matching performance for images in which
different identity pairs consisted of faces that were extremely similar to each other and

224

same identity pairs included naturally captured variation within a face (Chapter 2).
Celebrity lookalike images were used as naturally occurring imposter faces, and these
images allowed the creation of difficult different face pairs (one image of the celebrity
and one image of a lookalike for that celebrity) which appeared in a face-matching task
along with same identity pairings (two different [ambient] images of the same celebrity’s
face). Participants were instructed to make same or different identity judgments for each
of the image pairs. Experiment 1 demonstrated a graded effect of familiarity for the
lookalike task - unfamiliar viewers made many identity matching errors, whereas familiar
viewers performed with near perfect accuracy. My pattern of findings in terms of
familiarity and performance demonstrated lower accuracy on my celebrity lookalike task
(mean performance accuracy = 72%) than performance for unfamiliar viewers on the
GFMT, a standardised face-matching task, which contained cooperative stimuli (mean
performance accuracy long version = 89.9%, short version [hardest 40 items] = 81.2%). To
model the applied problem of varied image quality, Experiments 2 and 3 proceeded to
make the task harder still by reducing image quality through pixelation. The images were
degraded making them challenging but also realistic of the image type often acquired
from zoomed in digital images. A graded familiarity advantage survived through
Experiment 2, however for the highest pixelation level (Experiment 3) performance was
around chance for all but the extremely familiar viewers. These findings highlight firstly
that performance for challenging stimuli image matching is even worse than the poor
performance already established for cooperative face image matching. Additionally, the
findings highlight familiarity as a graded concept. Finally my findings suggest that the
familiarity advantage has limits – only extreme familiarity with a face could help in the
case of coarse images, and this advantage remained only for same identity trials
(Experiment 3).

I found that image quality affected overall face-matching performance, and the influence
that familiarity could have on performance. Matching accuracy was lower for degraded
(mid-pixelated and coarsely pixelated) versions of the celebrity and lookalike images
(Experiments 2 & 3) than for the un-manipulated (fine quality) versions of the images
(Experiment 1). In forensic investigations, the images that are available for comparison

225

often have this pixelated appearance. It was therefore important to investigate ways of
improving identification accuracy for these poor quality images (Chapter 3). Chapter 3
explored whether techniques that have been successful in improving cooperative facematching accuracy, could also improve accuracy for challenging images. Blurring the
pixelated images (Experiment 4), pooling judgments via crowd analysis (Experiment 5)
and superior ability of super-recognisers (Experiment 6) all improved face-matching
performance for challenging images. I also found that some of these techniques could be
combined for additional benefits.

I next looked at a naturally occurring image manipulation – change of camera-to-subject
distance – to investigate how this change affected facial appearance for individuals across
photographs, and whether such changes in appearance caused difficulties for facematching. Thus, this was an investigation of an unintentional change to appearance,
which may in turn influence identity judgment. Measuring face images (Experiment 7)
showed that changing camera-to-subject distance resulted in non-linear changes in
distances between features of a face, such that ratio measurements between features
were not preserved when camera-to-subject distance was altered between photographs.
Experiment 8 showed that these differences in facial configurations across images caused
matching difficulty for unfamiliar viewers. Performance for same identity pairs was
poorer when matching across images of varied camera-to-subject differences, compared
to performance for images taken from the same camera-to-subject distance. Familiar
viewers were unaffected by the camera-to-subject manipulation. I also showed that
viewers compensate for changes to camera-to-subject distance when distance cues are
available (Experiment 9), implying a high-level perceptual constancy for face shape.

The focus of investigation moved next to intentional appearance changes – deliberate
disguise. I created the FAÇADE image database as a resource that includes evasion and
impersonation disguise stimuli as well as undisguised comparison images of the same
identities. These images made it possible to perform a direct comparison between
disguised and undisguised face-matching performance. I found that deliberate disguise

226

impaired face-matching for unfamiliar viewers (Experiment 10), even when participants
were informed of the disguise manipulations (Experiment 11). Not all disguises caused
equal difficulty. Evasion impaired performance more than impersonation. Moreover
matching accuracy was higher when the target and model’s impersonations were based
on random matching than when they were of naturally similar appearance prior to the
impersonation disguise. Interestingly, familiar viewers performed better than unfamiliar
viewers overall (Experiment 12), but they too were worse at matching evasion faces than
undisguised face pairs. Chapter 6 was more exploratory, and investigated how the models
in the FAÇADE database disguised themselves and what made for an effective disguise. A
social inference experiment concluded the investigation to find out whether the disguise
manipulations applied affected the social inferences made for the individual. Viewers
made significantly different social trait inferences about the models, when they viewed
an undisguised image of the model’s face than when they viewed the same model’s face
in evasion disguise. However, for the impersonation scenario, social inference ratings
were not significantly more similar between images of the target face and impersonation
face, than between the undisguised model image and impersonation image. This suggests
that whereas social inference related changes are important for creating evasion disguise,
other factors are more important for creating impersonation disguise.

In summary, the results of this thesis show that face-matching performance can be
impaired both by incidental changes in appearance (Chapters 2 & 4) and by deliberate
changes in appearance (Chapter 5). There are several methods of improving this poor
performance (Chapter 3). In some instances natural solutions can be exploited, including
taking performance from those with high face recognition aptitude (super-recognisers,
Experiment 6) and high-level perceptual constancy can account for camera-to-subject
distance related face changes (Chapter 4). Familiarity improved performance throughout.
I explored what people did to disguise themselves, and what works. Disguise is more
complex than previous studies have allowed for. For example evasion and impersonation
result in different levels of difficulty and the disguises themselves were created using
different techniques and facial manipulations. Effective disguise involves more than the
simply the addition of props, but it is not the case that there is a simple disguise formula

227

which can be applied to all – some disguise manipulations and combinations work for
some individuals but not for others. Not all questions relating to disguise could be
answered within this work.

7.2 Relation to Previous Research
Previous research had already established that unfamiliar face-matching performance is
poor. All of my experiments support this, as I consistently found poor performance for
unfamiliar viewers (Chapter 2, Chapter 4, Chapter 5). My findings add to this past
research by showing that accuracy for challenging (but realistic) image conditions is lower
than for cooperative face-matching, something that was already known to be poor (e.g.
Burton et al. 1999). In support of previous findings, my results show much better
performance for familiar viewers than unfamiliar viewers in each of my face-matching
tasks (Chapter 2, Chapter 4, Chapter 5). I also found a graded familiarity effect for facematching performance similar to that demonstrated by Clutterbuck & Johnston (2001,
2003), and implemented a new familiarity scale to capture this concept (Experiment 1).
There are also some important differences between my findings and past familiarity
advantages. In past research familiarity has generally led to performance to be at ceiling
(e.g. Burton et al. 1999; Bruce et al. 2001), and this has made it difficult to quantify
changes in performance. The challenging nature of my tasks took familiar viewer
performance off ceiling level (Chapter 2, Chapter 5), revealing important gradations at the
end of the range. Furthermore, limits of familiarity began to become clear in Experiment
3, when familiarity was tested to destruction. The graded nature of the familiarity
advantage ceased to exist for the coarse image version of the lookalike task. However the
advantage did not break down completely; extremely familiar viewers outperformed the
unfamiliar viewers at the task, but this was only true for same person trials and may
reflect the greater variation of appearances held for familiar faces, accessed when the
identity of the target celebrity is recognised by the viewer. Familiar viewers performed
the same task with high levels of accuracy for both same and different face image pairs
when the image quality was good (Experiment 1) or slightly degraded (Experiment 2) and
past studies report that familiar viewers remain unaffected by image degradation (Burton

228

et al. 1999). These results thus suggest that the combination of the similarity of faces and
severely degraded image quality made the task difficult for familiar viewers, rather than
either of these factors taken in isolation.

Although I began to find limits of familiarity in Experiment 3, it was in Experiment 12 that
I saw the performance of familiar viewers suffer due solely to the model images being
challenging, rather than due to challenging images combined with degraded image
quality. In Experiment 12, I found that for evasion disguise faces, familiarity improved
performance compared to performance for unfamiliar viewers, but familiarity did not
completely compensate for the effect of disguise. Familiar viewers made more errors
when matching evasion disguise face pairs than undisguised versions of these pairs. The
results of Experiments 10, 11 & 12 demonstrate that the evasion of identity created a
more difficult face-matching scenario than impersonation of identity. The scenario in
Experiment 12 where familiar viewers performed with lower accuracy for disguised and
undisguised faces occurred for evasion disguise only. Indeed evidence from other
experiments in this thesis is consistent with the deliberate disguise findings. The celebrity
lookalike images presented in Experiment 1, for which familiar participants performed
with very few errors, was reflective of an impersonation scenario as lookalikes could be
considered as impersonators of the celebrity. Therefore results are agreeing that familiar
viewers are generally able to perform highly for matching faces that involve
impersonation (Experiments 1 & 12) but evasion disguise can cause familiar viewers
matching difficulty (Experiment 12).

Familiar viewers outperformed unfamiliar viewers in all tasks (including the evasion
condition described above). However there are many applied situations e.g. passport
security checks and monitoring of CCTV images, where it is not possible for viewers to be
familiar with the faces concerned. With this in mind, I was keen to investigate methods
that had previously been found to be effective in improving face-matching performance
and apply them to my tasks which involved more challenging stimuli in this thesis. I found
that both image manipulation (blurring of pixelated images) and crowd analysis could

229

improve accuracy for pixelated face-matching (Experiments 4 & 5). I also found that
super-recognisers made more accurate judgments than comparison observers. This result
demonstrates that super-recognisers’ superior face recognition ability extends beyond
good quality images and implies that their high performance does not rely solely on fine
scale information in face images. This finding therefore supports the notion of recruiting
super-recognisers for face recognition roles, and extends the past evidence for this
proposal by showing that super-recognisers also hold a superior ability for matching very
challenging image pairs as well as cooperative image pairs (e.g. White et al., 2015).
Additionally, for the specific case of improving performance on images with changed
camera-to-subject distance, providing accurate distance cues is a means of boosting
performance (Experiment 9).

In terms of past research on deliberate disguise, my findings support the most basic
finding of previous face recognition research:- disguise presence impaired face
recognition. None of the present findings go against those of previous studies, however
my research builds upon previous disguise research in several important ways. Past
research had studied disguise with a focus solely on evasion. A key finding from the
experiments presented here is that disguise cannot be understood as a unitary
manipulation. I found consistent differences between evasion and impersonation disguise
in terms of matching accuracy (higher accuracy for impersonation) and also ways in which
the disguises were executed by the models (higher relevance on external features for
impersonation). Additionally I showed that there is more to disguise than occlusion of
facial features. Past research relied almost entirely on occlusion manipulations such as
glasses and facial hair. Those manipulations tell us rather little about disguise. Any form
of occlusion is apt to reduce matching accuracy as it reduces the available information.
Moreover occlusion also would not be an effective disguise in many security scenarios, as
the props that occluded features may have to be removed for identity decisions to be
reached. Here I have shown that disguise can impair face recognition even without
occluding features. In addition to this, I highlighted the importance of free disguise. By
giving participants the freedom to disguise themselves as they wished, rather than as
prescribed by the experimenters, I was able to show that people naturally use many

230

different methods to disguise themselves and that some methods are more effective than
others. Models did far more than simply add props in creating their disguise
manipulations. Investigation of what made a successful disguise showed that there was
no simple disguise recipe. Contrary to the implicit assumption of previous studies,
approaches to disguise are rather idiosyncratic - some disguise manipulations (e.g.
manipulating hairstyle, copying or changing expression) were effective in the case of
some identities but not for others. Overall, effective Evasion involved more non-feature
based changes, whereas effective impersonation was created through internal and
external featural manipulations.

7.3 Theoretical Implications
Up until now this discussion has focussed upon the practical advances gained from my
research, realistic challenging images, performance enhancements and insights into
deliberate disguise. The experiments that I conducted also resulted in implications for
theory. Throughout the thesis I argued that my findings support a theoretical standpoint
on face learning that has within person variability at its heart: what viewers are learning
when they become familiar with a face is all of the different ways which that face can look
(Jenkins et al., 2011). The range of possible appearance can be constructed as a form of
face space that is specific for that identity. Figure 7.2 depicts this face space as a
multidimensional volume, encapsulating the range of accepted appearances for that face.
Any face image that falls inside the volume is accepted as that individual, and any image
that falls outwith this volume is not. Exposure to a person’s face helps viewers to refine
the face space for that individual. Less familiar viewers also hold their face space
representations of individuals but their face space is less refined as a result of their
limited exposure to the face and limited experience of appearances that the face can
take, making unfamiliar viewers more likely to make identity judgment errors. Face space
is refined as exposure increases, therefore refinement of face space reflects the graded
familiarity advantage found in Experiments 1 & 2. I found that performance on my
celebrity lookalike task improved as a result of increasing familiarity of the faces
concerned. For viewers who were unfamiliar with the celebrity faces, the lookalike face

231

image likely fell within the accepted range of faces for the celebrity leading to an
incorrect identity match decision. As familiarity increased, less of these errors occurred,
with the lookalike image more often falling outside the real celebrity’s face space.

My findings also suggest arguments against both the configural and featural account of
face recognition. Whilst configural and featural details of a face may be used to aid face
recognition in some ways, I suggest that neither configural nor featural information is the
key to face recognition. Experiment 7 showed that configurations of a face, as
represented in images, are not constant (e.g. Kleinberg et al. 2007). In Experiment 7,
distances between features underwent non-linear changes as a result of change in
camera-to-subject distance. But only unfamiliar viewers were affected by the distance
manipulation. Familiar viewers performed highly no matter the camera-to-subject
distance, supporting the theory that familiar viewers have expertise of all the different
ways that a face can look. If they have previously seen the face across a range of distance
and image conditions, then photographs of the same person taken from different
distances would still fall into the accepted range of faces that a familiar viewer holds for
the face, but outside the accepted range of faces held by an unfamiliar viewer.

Figure 7.1 Schematic representations of the disguise manipulations with regards to face space. Each bubble
represents one individual’s face space.

232

The theoretical standing also fits alongside the experimental results for disguise. Evasion
disguise led to more errors than impersonation disguise. With relation to face space,
there are more ways that a person can leave their own face space (evade identity), than
enter someone else’s (impersonate identity), for example changing hair colour may make
someone look unlike themselves, but changing hair colour would only help for
impersonation if the hair colour was changed to match that of the person that you want
to look like. Impersonation involves moving appearance outwith the accepted face space
for your own face, and into the accepted space of someone else. Impersonation change is
thus limited to one direction (changes that take you towards, and ideally into, the face
space of the person who is to be impersonated) whereas evasion can involve change in
any direction as long as the face is removed out of its own face space (see Figure 7.1).

Exploration and Experiment 13 in Chapter 6 suggested a more complex component to
face space with relation to disguise. Many different techniques were used by my models,
and in various combinations, to create the various evasion and impersonation disguises.
Some disguises worked for some identities but not so well for others. It is possible that
certain faces were more unique or generic, hence making it easier or harder for the
models to move out of their own face space or into another’s. Making changes to a face,
which lead to a change in social inference, was found to be an effective way of moving a
face outwith its own face shape in the case of evasion disguise.

Evasion disguise was linked to change in social inference whereas impersonation disguise
was not. Familiar viewers were impaired only by evasion of identity, not impersonation
disguise. Notably this is the only disguise change that was linked to significant changes in
inferred social inferences. Familiar viewers had an additional advantage for identifying
impersonation faces over evasion faces. In impersonation trials the familiar viewers were
familiar with both the impersonator and the person being impersonated – therefore
there were two ways that they could approach the matching task. The viewer could either
perceive the true identity of the impersonator, or decide that the impersonator falls
outside the accepted range for the person being impersonated. In future studies it would

233

be possible to dissociate these possibilities by conducting a study in which viewers were
familiar with only one of the faces in each pair. In any case it is clear that impersonation is
harder to accomplish than evasion because the direction of change is constrained.

What is not yet known is exactly what it is that people are learning in a face when they
are becoming familiar with it. Face space is proposed as a multidimensional space. An
identity may change in appearance across many of these dimensions, but for recognition
to occur the match must presumably fall within the expected range for at least one of the
dimensions. I have talked about face space in a rather abstract representation as
described by Burton et al. (2015). Dimensions that are relatively unaffected by disguise,
could be identified by applying Principal Components Analysis (PCA) to multiple images
(e.g. 20 undisguised images) of the model in order to establish a face space for the
identity. New images of the identity could then be entered into this face space along with
the disguised images to see where each of these images fall. This would show if there are
any dimensions that remain constant within the disguise face images, and also perhaps
reveal which dimensions of change are necessary in the creation of effective evasion and
impersonation disguises. As it stands it is unclear what it is that people are recognizing
when they are successfully seeing through disguise to recognizing a face. It may be that
critical information is the same for every case of disguise, although it is likely to be
different for evasion and impersonation. This interpretation is based on the differences
found between these types of disguises so far. It is also possible that different things are
used to see through each disguise created by each model because the underlying face of
each of these models is different to begin with. This type of investigation for disguise
could also aid the understanding of undisguised (normal) face recognition. Whatever
familiar viewers can ‘see through’ in the case of disguise cannot be what is critical for
identification.

234

7.4 Practical Implications
Several practical implications follow from my research. First of all there is a need to
acknowledge that face-matching involving challenging images is less reliable than
cooperative face-matching. Moreover, merely acknowledging this problem is not enough.
In a striking demonstration of this (Experiment 11), being aware that a face may be
disguised did not improve matching performance. Instead specific methods such as
blurring of pixelated images (Experiment 4), crowd analysis of identity judgments
(Experiment 5) and personnel selection (Experiment 6) will likely be required. Even then
errors will not be eliminated entirely.

Of all the manipulations in this thesis, familiarity improved performance most. Familiar
viewers are thus recommended to make identity judgments for similar challenging images
in forensic situations. As the graded effects in Experiments 1-4 emphasise, a little
familiarity with a face is better than none at all, and a lot is better than a little. In
summary, the more familiar a person is with the face or faces concerned, the more likely
the correct identity judgment will be reached.

In Experiment 8, I found that photographic conditions – specifically camera-to-subject
distance – could impair identity judgments for unfamiliar faces. My Experiments 2 & 3
show a similar decrease in performance for pixelated faces. However in Experiment 8, I
confirmed that changing the distance from which a photograph is taken from can make it
difficult to compare identity across images. This finding highlights that using photographs
for identity comparisons is problematic even when image quality is high. Distance cues
can aid recognition when these are available, but cues are not readily available from most
photograph images, especially images used for identity confirmation, as these images are
normally cropped around the face, removing any background related cues that may have
indicated subject to camera distance. I found that images are more accurately matched
when taken from the same distance. A practical implication from this would be that
where possible, consistency should be applied when photographing individuals in security

235

and forensic scenarios. For example, police could photograph all suspects from the same
distance across all photos of an individual. This standardisation would make it easier to
identify the same person across multiple images and also make it less likely that a suspect
would be mistaken for another person in a side-by-side image comparison scenario.
Camera-to-subject distance changes would however remain problematic for archived
images and some CCTV footage, but introducing standardisation to station captured
images may aid some identity scenarios. Using familiar viewers to identify faces taken of
different distances will largely overcome the problem as they are far less affected by
camera-to-subject distance change.

My disguise research has implications for applied settings. I show that people are very
poor at matching disguised face images, suggesting that it is rather easy to carry out a
successful disguise. Disguises that involve evading identity are more likely to attract
errors than disguises that involve the impersonation of somebody else. There are
practical implications here in terms of both criminal cases of identity fraud and also
undercover policing. In terms of criminal disguise activity, my research highlights that
cases of impersonation and evasion may go undetected. Successful evasion disguise is
particularly concerning as disguise could make it particularly difficult to catch a suspect.
Impersonation disguise carries many security threats in terms of identity fraud, which
could have catastrophic results particularly in terms of allowing a dangerous or
unauthorised person into a country or providing them with access to information, which
their true identity should not be granted. Disguise is problematic for face recognition, and
errors can pose danger. Therefore steps discussed above should be taken to try and
minimise fraud and identity evasion through disguise. The disguise results do however
suggest that undercover police can successfully keep their true identity hidden by using
evasion disguise.

There are also some important implications from this thesis in relation to the
experimental psychology research practice for the study of face recognition. First my
findings reiterate the importance of a distinction between familiar and unfamiliar face

236

matching, and extend this to provide further evidence for a graded familiarity advantage
in face recognition. Past studies have tested face-matching performance using
cooperative face-matching stimuli. I suggest that measuring performance for cooperative
stimuli images does not capture face-matching performance levels for uncooperative
stimuli. I quantified the performance costs for various challenging situations and found
lower accuracy levels for the images that I tested, than those reported by previous
studies. It is important to study uncooperative image performance, especially when face
recognition performance is so closely linked to many security systems or included in
witness testimonials for crimes. Suspects may deliberately make their identification effort
difficult. I also highlight important distinctions within disguise. The previous research on
disguise was based on evasion. I have extended upon this in two ways; i) by investigating
evasion disguise as being more than just the addition of prescribed props which occlude
facial features, and ii) by looking at impersonation which had been completely ignored by
past studies of disguise. Individuals spontaneously use many different types of
manipulations to disguise themselves, and these disguises impair identity performance
even when there is no occlusion of internal facial features. My research has found clear
distinctions between evasion and impersonation, both in terms of face-matching
performance and the kinds of disguise manipulations applied. Future research will have to
acknowledge these distinctions if we are to arrive at a complete understanding of
deliberate disguise.

7.5 Future Directions
This thesis expands the research for face recognition in challenging situations, however
there are areas of investigation that I believe warrant future study. Firstly, it would be
interesting to test face memory performance for the FAÇADE database images. I found
poorer performance accuracy for matching disguised faces than undisguised faces.
Matching faces is presumably an easier task than face recognition tasks that involve a
memory component (Megreya & Burton, 2008). Previous memory studies have tested for
recognition of disguised faces in direct comparison to performance for memory of
identical images (Patterson & Baddeley, 1977). My FAÇADE database includes different

237

images of the same faces undisguised and in disguise, and evasion and impersonation
disguise images. This database would thus allow a more thorough and experimentally
sound test of face memory performance for disguised versus undisguised faces, and also
allow performance for evasion disguise to be compared with impersonation disguise.
Performance could be even worse for disguise in a memory task, as in the matching task
both images could be compared in side-by-side comparison. It is not obvious whether the
differences in performance found for evasion and impersonation would remain in a
memory scenario. I believe this to a be a particularly interesting comparison, as
impersonation could trick viewers to incorrectly ‘remember’ the real target face and
reject an image presented of the true identity undisguised. There are many reports of
erroneous eyewitness memory for studies that do not include impersonation disguise
(Bruce, 1988), therefore impersonation disguise may increase error rates. Additionally,
evasion disguise may result in rejection of the true identity face presented in an
undisguised form.

It would also be interesting to test whether super-recognisers may also be able to
improve performance accuracy in the case of disguise images. This would need to be
confirmed by future research. If super-recognisers do perform better than our unfamiliar,
or even possibly familiar controls, then super-recognisers could be of great help in
identity efforts involving disguise. It would be important to explore whether superrecognisers or familiar viewers provide a more accurate viewer group to call on for
making identity decisions. Based on the only directly comparable data in this thesis for an
unfamiliar group of super-recognisers (Experiment 6 [mean performance accuracy for this
group = 76%]) and the highest familiarity group of comparison participants (Experiment 2
[mean performance accuracy for this group = 76%]), I would predict that unfamiliar superrecognisers and familiar comparison viewers are equally good at matching faces.

Another interesting investigation for the future research would be to explore the effect of
methods of improving performance (Experiments 4, 5 and 6) in the combinations that
were not addressed in this thesis. I found that when I tested methods in combination, this

238

led to further improvements in performance than any of the methods used alone. Due to
the limited opportunity to work with super-recognisers I was unable to test all of these
combinations. Specifically, it is currently unknown whether blurring the pixelated images
would have further improved the performance of super-recognisers, and also whether
combining all three methods could result in improvements greater than those from
combining any two methods.

A key future experiment will be the Principal Component Analysis (PCA) of disguised faces
outlined in the theoretical advances section in this discussion. The proposed experiment
will allow theoretical advancement in terms of what it is that allows viewers to see
through disguise. This has particular relevance to the theory of face space, and the
refinement of face space with familiarity. The familiarity component of this theory can be
explored through manipulating the number of reference face images entered into the
PCA. This will help to establish what it is about a face which stays constant in disguise,
and whether there are reliably constant factors for it at all, or if factors differ by identity.

Finally, I believe it would be interesting to investigate machine performance for disguised
faces. Computer algorithms are now being used in many security situations including
passport security, but as with human performance, machine performance has generally
been tested using cooperative stimuli. Attempts to test machine performance for
disguised faces have many of the same limitations as previous human disguise
investigations, namely testing has been limited to evasion disguise where images were
disguised through occlusion of features. The props which cause occlusion disguise would
be removed in most security scenarios, therefore testing machine performance on these
images is of limited use for security algorithms used in passport control. Real cases of
identity fraud more often include impersonation or the morphing of a new holder’s face
with the face of the true passport owner. Impersonation is therefore particularly relevant
in terms of security related identity fraud, but has not yet been tested in terms of
machine performance. My FAÇADE database provides stimuli void of occlusion of
features, and includes evasion and impersonation disguise. Future studies could therefore

239

use this database to more fully explore machine performance for disguised faces and
compare this performance for undisguised versions of the same faces. As data has already
been collected for both unfamiliar and familiar human viewers, machine performance
could then be directly compared to performance for each of these human viewer groups,
helping to answer the question of whether humans or machine are better at matching
disguised faces. Greater performance accuracy could perhaps be achieved through fusing
the performance of humans and machine, this has been a successful method of improving
face recognition accuracy in past studies (O’Toole et al. 2007).

In summary, previous estimates of identification accuracy are likely to be overestimates if
the people who were being identified were cooperating with their identity effort. Applied
cases of face recognition will more likely include images with incidental or deliberate
differences to their own appearance or similarities with another person, as have been
investigated by this thesis. Familiar viewers are generally considered to be exceptionally
good at identifying faces and this thesis has shown that this familiarity advantage indeed
extends to performance for both incidental and deliberate appearance change. However,
challenging situations start to reveal limits of this advantage. For example, familiar
viewers’ face-matching performance is impaired by evasion disguise – this raises
interesting questions about what appearance change a familiar viewer will allow to
classify an individual as still being the same person. The results of this thesis suggest that
familiar viewers are better than unfamiliar viewers for the experience of a face that they
have learnt. There are also many applied implications of this research. For example,
identity situations in the real world are often linked to impersonation disguise, however
there has been a great deal of disconnect between lab experiments and the real world
problem of identity fraud, as impersonation had not been addressed by previous lab
experiments. I found that impersonation is harder to pull off than evasion disguise. Poor
performance for matching evasion disguise faces is problematic for capturing criminals or
missing people on police watch lists, but suggests that evading identity of undercover
police officers will likely be effective. There is still a lot more to learn about face
recognition in challenging situations, especially for the case of disguise, and I hope the
FAÇADE database will continue to aid this investigation.

240

References
Alain, C., & Proteau, L. (1980). Decision making in sport. In C.H. Nadeau, K.M. Halliwell,
K.M. Newell, & G.C. Roberts (Eds.), Psychology of motor behaviour and sport. 465-477.
Champaign, IL: Human Kinetics.
Andrews, T. J., & Ewbank, M. P. (2004). Distinct representations for facial identity and
changeable aspects of faces in the human temporal lobe. NeuroImage, 23(3), 905–913.
doi:10.1016/j.neuroimage.2004.07.060
Aristotle, Jowett, B., & Davis, H. W. C. (1920). Aristotle's Politics. Oxford: At the Clarendon
Press.
Bachmann, T. (1991). Identification of spatially quantised tachistoscopic images of faces:
How many pixels does it take to carry identity? European Journal of Cognitive Psychology,
3, 85-103.
Behrmann M, Avidan G 2005 Congenital prosopagnosia: face-blind from birth. Trends
Cogn. Sci. 9, 180–187
Bindemann, M., Attard, J., Leach, A. M. Y., & Johnston, R. A. (2013). The Effect of Image
Pixelation on Unfamiliar-Face Matching, 717(November), 707–717.
Bindemann, M., Burton, A.M., Leuthold, H., & Schweinberger, S.R. (2008). Brain potential
correlates of face recognition: Geometric distortions and the N250r brain response to
stimulus repetitions. Pscyhophysiology,45(4), 535-544.
Bobak, A. K., Dowsett, A. J., & Bate, S. (2016). Solving the border control problem:
Evidence of enhanced face matching in individuals with extraordinary face recognition
skills. PLoS ONE, 11(2), 1–13. doi:10.1371/journal.pone.0148148
Bonner, L., Burton, A.M. & Bruce, V. (2003). Getting to know you: how we learn new
faces. Visual Cognition. 10, 527-536.

241

Borkenau, P. Brecke, S. Mottig, C., & Paelaecke, M. (2009). Trait inferences: Sources of
validity at zero acquaintance. Journal of Personality and Social Psychology, 62, 645-657.
Bradshaw, J.L., & Wallace, G. (1971). Models for the processing and identification of faes.
Perception and Psychophysics, 9, 443-448.
Brigham, J.C.& Bothwell, R.K. (1983). The ability of prospective juroros to estimate the
accuracy of eyewitness identification. Law and Human Behaviour. 7, 19-30.
Bruce, V. (1982). Changing faces: visual and non-visual coding in face recognition. British
Journal of Psychology. 73, 105-116.
Bruce, V. (1986). Influences of familiarity on the processing of faces. Perception, 15(4),
387–97. Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/3822723
Bruce, V. (1988) Recognising Faces. London: Erlbaum.
Bruce, V. (1994). Stability from variation: The case of face recognition the M.D. Vernon
memorial lecture. The Quarterly Journey of Experimental Psychology Section A: Human
Experimental Psychology/abstract content. 47(1), 5-28.
Bruce, V., Doyle, T., Dench, N. & Burton, M. (1991). Remembering facial configurations.
Cognition, 38(2), 109-144.
Bruce, V., Henderson, Z, Greenwood, K., Hancock, P.J.B. & Burton, A.M. (1999).
Verification of face identities from images captured on video. Journal of Experimental
Psychology – Applied, 5, 339-360.
Bruce, V., Henderson, Z., Newman, C., Burton, A.M. (2001). Matching Identities of
Familiar and Unfamiliar Faces Caught on CCTV. Journal of Experimental Psychology:
Applied, 7, 207-218.
Bruce, V. & Langton, S. (1994). The use of pigmentation and shading information in
recognising the sex and identity of faces. Perception, 23, 803-822.

242

Bruyer, R. & Coget, M. C. (1987). Features of laterally displayed faces: saliency or top
down processing? Acta Psychologica, 66, 103-114.
Bryan, R., Perona, P. & Adolphs, R. (2012). Perspective Distortions from Interpersonal
Distance Is an Implicit Visual Cue for Social Judgments of Faces. PLoS ONE 7(9): e45301.
doi:10.1371/journal.pone.0045301
Burton, a. M., Kramer, R. S. S., Ritchie, K. L., & Jenkins, R. (2015). Identity From Variation:
Representations of Faces Derived From Multiple Instances. Cognitive Science, 40, 1–22.
doi:10.1111/cogs.12231
Burton, a. M., Schweinberger, S. R., Jenkins, R., & Kaufmann, J. M. (2015). Arguments
Against a Configural Processing Account of Familiar Face Recognition. Perspectives on
Psychological Science, 10(4), 482–496. doi:10.1177/1745691615583129
Burton, a. M., & Vokey, J. R. (1998). The Face-Space Typicality Paradox: Understanding
the Face-Space Metaphor. The Quarterly Journal of Experimental Psychology Section A,
51(3), 475–483. doi:10.1080/713755768
Burton, A. M., White, D., & McNeil, A. (2010). The Glasgow Face-matching Test. Behaviour
Research Methods, 42(1), 286-291.
Burton, A. M., Wilson, S., Cowan, M., Bruce, V. (1999) Face recognition in poor-quality
video: Evidence from security surveillance. Psychol Sci 10: 243–248. doi: 10.1111/14679280.00144
Clutterbuck, R., & Johnston, R. a. (2002). Exploring levels of face familiarity by using an
indirect face-matching measure. Perception, 31(8), 985–994. doi:10.1068/p3335
Clutterbuck, R., & Johnston, R. a. (2004). Demonstrating the acquired familiarity of faces
by using a gender-decision task. Perception, 33(2), 159–168.
Clutterbuck, R., & Johnston, R. a. (2005). Demonstrating how unfamiliar faces become
familiar using a face matching task. European Journal of Cognitive Psychology, 17(1), 97–
116. doi:10.1080/09541440340000439

243

Coates, Tim (1999). The strange story of Adolph Beck. London: Stationery Office.
Collishaw, S. M., & Hole, G. J. (2000). Featural and configurational processes in the
recognition of faces of different familiarity. Perception, 29(8), 893–909.
doi:10.1068/p2949
Costen, N. P., Parker, D. M.m & Craw, I. (1994). Spatial content and spatial quantization
effects in face recognition. Perception, 23, 129-146.
Costen, N. P., Parker, D. M., & Craw, I. (1996). Effects of high-pass and low-pass spatial
filtering on face identification. Perception and Psychophysics, 58, 602-612.
Davies, G., Ellis, H. D., & Shepherd, J. (1978). Face recognition accuracy as a function of
mode of representation. Journal of Applied Psychology, 63(2), 180–187.
doi:10.1037//0021-9010.63.2.180
Dhamecha, T. I., Singh, R., Vasta, M. & Kumar, A. (2014). Recognising disguised faces:
Human and machine evaluation. PlosOne.9(7): e99212.
Diamond, R. & Carey, S. (1986). Why faces are and are not special: an effect of expertise.
Journal of Experimental Psychology: General. 115, 107-117.
Donders, F. C. (1868/1969). On Speed of Mental Processes. Acta Psychologica, 30, 412431.
Dowsett, A. J., & Burton, a. M. (2014). Unfamiliar face matching: Pairs out-perform
individuals and provide a route to training. British Journal of Psychology, n/a–n/a.
doi:10.1111/bjop.12103
Duchaine, B. (2011). Developmental Prosopagnosia: Cognitive, Neural and Developmental
Investigations. In A. J. Calder., G. Rhodes, M. H. Johnson, & J. V. Haxby (Eds.), The Oxford
Handbook of Face Perception (pp.821-838). Oxford, UK: Oxford University Press.
Duchaine, B., & Nakayama, K. (2006). The Cambridge Face Memory Test: Results for
neurologically intact individuals and an investigation of its validity using inverted face

244

stimuli

and

prosopagnosic

participants.

Neuropsychologia,

44(4),

576–585.

doi:10.1016/j.neuropsychologia.2005.07.001
Ellis, H.D., Shepherd, J.W. & Davies, G.M. (1979). Identification of familiar and unfamiliar
faces from internal and external features: some implications for theories of face
recognition. Perception, 8, 431-439.
Flow, H & Ebbesen, B. (2007). The effect of lineup member similarity on recognition
accuracy in simultaneous and sequential lineups. Law and Human Behaviour. 31(1), 3352.
Galper, R.E. (1970) Recognition of faces in photographic negative. Psychonomic
Science,19, 207-208.
Galton, F. (1907). Vox populi - The wisdom of crowds. Nature, 75, 450–451.
doi:10.1038/075450a0
Ghent, L. (1960). Recognition by children of realistic figures presented in various
orientations. Canadian Journal of Psychology, 14, 249-256.
Golstein, A.G. (1965). Learning of inverted and normally orientated faces in children and
adults. Pscyhonomic Science, 3, 447-448.
Grill-Spector, K., Kushnir, T., Edelman, S., Avidan, G., Itzchak, Y., & Malach, R. (1999).
Differential processing of objects under various viewing conditions in the human lateral
occipital complex. Neuron, 24(1), 187–203. doi:10.1016/S0896-6273(00)80832-6
Haig, N. D. (1984). Faces in Perception. Perception, 13, 1158–1165.
Hancock, P. J. B., Bruce, V., & Burton, A. M. (2000). Recognition of unfamiliar faces, 4(9),
330–337.
Harmon, L. & Julesz, B. (1973). Masking in Visual Recognition: Effects of Two-Dimensional
Filtered Noise. Science. New Series, 180. 1194-1197.

245

Harper, B. & Latto, R. (2001). Cyclopean vision, size estimation, and presence in
orthostereoscopic images. Presebce, 10(3), 312-330.
Henderson, Z., Bruce, V., & Burton, a. M. (2001). Matching the faces of robbers captured
on video. Applied Cognitive Psychology, 15(4), 445–464. doi:10.1002/acp.718
Henle, M. (1942). An experimental investigation of past experience as a determinant of
visual perception. Journal of Experimental Psychology, 30, 1-22.
Hill, H., & Bruce, V. (1996). Effects of lighting on the perception of facial surfaces. Journal
of Experimental Psychology. Human Perception and Performance, 22(4), 986.
doi:10.1037/0096-1523.22.4.986
Hochberg, J. & Galper, R.E. (1967). Recognition of faces: An exploratory study.
Psychonomic Science, 9, 619-620.
Hole, G. J., George, P. A., Eaves, K., & Rasek, A. (2002). Effects of geometric distortions on
face-recognition performance. Perception. doi:10.1068/p3252
Home

Office.

(2012).

False

ID

Guidance.

The

Home

Office.

https://www.gov.uk/government/uploads/system/uploads/attachment_data/file/98108/
false-id-guidance.pdf
Huff, CR. (1987). Wrongful Conviction: Societal tolerance of injustice. Research in social
problems and public policy. 4, 99-115.
Jenkins, R., & Burton, A. M. (2011). Stable face representations. Philosophical
Transactions of the Royal Society of London. Series B, Biological Sciences, 366(1571),
1671–1683. doi:10.1098/rstb.2010.0379
Jenkins, R., & Kerr, C. (2013). Identifiable images of bystanders extracted from corneal
reflections. PLoS ONE, 8(12), 8–12. doi:10.1371/journal.pone.0083325
Jenkins, R., White, D., Van Montfort, X., & Burton, A.M. (2011). Variability in photos of the
same face. Cognition, 121(3), 313–23. doi:10.1016/j.cognition.2011.08.001

246

Johnston, A. Hill, H. & Carman, N. (1992). Recognising faces: effects of lighting direction,
inversion and brightness reveral. Perception, 21, 365-375.
Kanade, T. (1977). Computer recognition of human faces. Stuttgart: Birkhauser Verlag.
Kastner, S., Pinsk, M. A., De Weerd, P., Desimone, R., & Ungerleider, L. G. (1999).
Increased activity in human visual cortex during directed attention in the absence of
visual stimulation. Neuron, 22(4), 751–761. doi:10.1016/S0896-6273(00)80734-5
Kelly, M.D. (1970). Visual identification of people by computer. Tech. rep. AI-130,
Stanford AI Project, Stanford, CA.
Kemp, R. I., Towell, N. & Pike, G. (1997). When seeing should not be believing:
Photographs, credit cards and fraud. Applied Cognitive Psychology, 11,211–222.
Kemp, R., Pike, G., White, P., & Musselman, a. (1996). Perception and recognition of
normal and negative faces: the role of shape from shading and pigmentation cues.
Perception, 25(1), 37–52.
Kenny, D. a., Horner, C., Kashy, D. a, & Chu, L. C. (1992). Consensus at zero acquaintance:
replication, behavioral cues, and stability. Journal of Personality and Social Psychology,
62(1), 88–97. doi:10.1037/0022-3514.62.1.88
King, a. J., Cheng, L., Starke, S. D., & Myatt, J. P. (2012). Is the true “wisdom of the crowd”
to copy successful individuals? Biology Letters, 8(2), 197–200. doi:10.1098/rsbl.2011.0795
Kleinberg, K. F., Vanezis, P., & Burton, a M. (2007). Failure of anthropometry as a facial
identification technique using high-quality photographs. Journal of Forensic Sciences,
52(4), 779–83. doi:10.1111/j.1556-4029.2007.00458.x
Krause, S., James, R., Faria, J. J., Ruxton, G. D., & Krause, J. (2011). Swarm intelligence in
humans: Diversity can trump ability. Animal Behaviour, 81(5), 941–948.
doi:10.1016/j.anbehav.2010.12.018

247

Lander, K., Bruce, V., & Hill, H. (2001). Evaluating the effectiveness of pixelation and
blurring on masking the identity of familiar faces. Applied Cognitive Psychology, 15, 101116.
Leder, H., & Carbon, C.C. (2006). Face-specific configural processing of relational
information. British Journal of Psychology, 97, 19-29.
Leikas, S., Verkasalo, M., & Lönnqvist, J. E. (2013). Posing personality: Is it possible to
enact the Big Five traits in photographs? Journal of Research in Personality, 47(1), 15–21.
doi:10.1016/j.jrp.2012.10.012
Light, L.L.; Kayra-Stuart, F. & Hollander, S. (1979) Recognition memory for typical and
unusual faces. Journal of Experimental Psychology: Human Learning and Memory. Vol
5(3), 212-228. http://dx.doi.org/10.1037/0278-7393.5.3.212

Liu, C.H. (2003). Is Face Recognition in Pictures Affected by the Center of Projection? IEE
International Workshop on Analysis and Modeling of Faces and Gestures, (Los Alamitos,
CA: IEEE Computer Society), 53-59.
Loftus, E. & Doyle, J. (1992). Eyewitness testimony: Civil and Criminal. Charlottesville, VA,
Lexis Law.
Matinez, A.M & Benavente, R (1998). The AR Face Database. CVC Technical Report.
Number 24.
Maurer, D., Grand, R. Le, & Mondloch, C. J. (2002). The many faces of configural
processing. Trends in Cognitive Sciences, 6(6), 255–260. Retrieved from
http://www.ncbi.nlm.nih.gov/pubmed/12039607
Megreya A.M., Burton A.M. (2007) Hits and false positives in face-matching: A familiaritybased dissociation. Perception & Psychophysics. 69, 1175–1184. doi: 10.3758/bf03193954
Megreya, A. M., Burton, a. M. (2006). Unfamiliar faces are not faces: Evidence from a
matching task. Memory & Cognition, 34 (4), 865–876.

248

Megreya, A.M., & Burton, a.M. (2008). Matching faces to photographs. Poor performance
in eyewitness memory (without the memory). Journal of Experimental Psychology:
Applied, 14(4), 364-372.
Megreya, A.M. & Bindemann, M. (2009). Revisiting the processing of internal and external
features of unfamiliar faces : The headscarf effect. Perception, 38, 1831-1848.
Menon, A., White, D. & Kemp. R. (2015). Variation in Photos of the Same Face Drives
Improvements in Identity Verification. Perception. doi: 10.1177/0301006615599902
Morrone, M, C., Burr, D, C., Ross, J. (1983). Added noise restores recognisability of coarse
quantized images. Nature, 305, 226-228.
O’Donnell, C., & Bruce, V. (2001). Familiarisation with faces selectively enhances
sensitivity to changes made to the eyes. Perception, 30(6), 755–764. doi:10.1068/p3027
O’Toole, A.J., Edelman, S. & Bulthoff, H.H. (1998). Stimulus specific effects in face
recognition over changes in viewpoint. Vision Research. 38, 2351-2363.
Patterson, K.E. & Baddeley, A.D. (1977). When Face Recognition Fails. Journal of
Experimental Psyhcology: Human Learning and Memory, 3(4), 406-417.
Pirenne, M. H. (1970) Optics, Painting, & Photography. New York: Cambridge University
Press.
Pizlo, Z. (1994). A theory of shape constancy based on perspective invarients. Vision
Research. 34(12). 1637-1658.
Ramanathan, N., Chowdhury, A.R. & Chellappa, R. (2004). Facial Similarity Across Age,
Disguise, Illumination and Pose. Proceedings of International Conference on Image
Processing, 3, 1999-2002.
Rhodes, G., Brennan, S. & Carey, S. (1987). Identification and ratings of caricatures:
Implications for mental representations of faces. Cognitive Psychology. 19, 473-497.

249

Righi, G., Peissig, J.J. & Tarr, M.J. (2002). Recognising disguised faces. Visual Cognition.
20(2), 143-169.
Robbins, R., & Mckone, E. (2003). Can holistic processing be learned for inverted faces?
Cognition, 88, 79–107. doi:10.1016/S0
Robertson, D. J., Noyes, E., Dowsett, A., Jenkins, R., Burton, A. M., & Burton, M. (2016).
Face recognition by Metropolitan Police Super-recognisers. PLoS ONE, 11(2): e0150036.
doi:10.1371/journal.pone.0150036
Rossion, B. (2009). Distinguishing the cause and consequence of face inversion: the
perceptual

field

hypothesis.

Acta

Psychologica,

132(3),

300–12.

doi:10.1016/j.actpsy.2009.08.002
Rule, N. O., & Ambady, N. (2011). Judgments of Power From College Yearbook Photos and
Later Career Success. Social Psychological and Personality Science, 2(2), 154–158.
doi:10.1177/1948550610385473
Russell, R., Duchaine, B., & Nakayama, K. (2009). Super-recognizers: people with
extraordinary face recognition ability. Psychonomic Bulletin & Review, 16(2), 252–257.
doi:10.3758/PBR.16.2.252
Sandford, A., & Burton, a. M. (2014). Tolerance for distorted faces: Challenges to a
configural processing account of familiar face recognition. Cognition, 132(3), 262–268.
doi:10.1016/j.cognition.2014.04.005
Sergent, J. (1986). Microgenesis of face perception. In Aspects of Face Processing (Ellis, H.
D. et al., eds). 17-33, Martinus Nijhoff.
Singh, R., Vatsa, M. & Noore, A. (2009). Face Recognition with Disguise and Single Gallery
Images. Image and Vision Computing, 27(3), 245-257.

250

Smith, E.E. & Nielsen, G.D. (1970). Representations and retrieval processes in short-term
memory: Recognition and recall of faces. Journal of Experimental Psychology, 85, 397405.
Stevenage, S. V. (1998). Which twin are you? A demonstration of induced categorical
perception of identical twin faces. British Journal of Psychology, 89(1), 39–57.
doi:10.1111/j.2044-8295.1998.tb02672.x
Sutherland, C. a M., Oldmeadow, J. a, Santos, I. M., Towler, J., Michael Burt, D., & Young,
A. W. (2013). Social inferences from faces: ambient images generate a three-dimensional
model. Cognition, 127(1), 105–18. doi:10.1016/j.cognition.2012.12.001
Tanaka, J. W. & Farah, M.J. (1993). Parts and wholes in face recognition. The Quarterly
Journal of Experimental Psychology, 46, 225-245.
Terry, R.L. (1993). How wearing eye glasses effects facial recognition. Current Psychology.
12(2). 151-162. 483-492.
Terry, R.L. (1994). Effects of facial transformations on accuracy of face recognition. The
Journal of Social Psychology. 134(4), 151-162.
Thomson, D.M. (1986). Face recognition: More than a feeling of familiarity? In H.D. Ellis,
M.A. Jeeves, F. Newcombe & A. Young (Eds.), Aspects of face processing. Dordrecht:
Matinus Nijhoff.
Todorov, A., Said, C. P., Engell, A. D., & Oosterhof, N. N. (2008). Understanding evaluation
of faces on social dimensions, (October), 455–460. doi:10.1016/j.tics.2008.10.001
Toseeb, U., Keeble, D. R. T., & Bryant, E. J. (2012). The significance of hair for face
recognition. PLoS ONE, 7(3), 1–8. doi:10.1371/journal.pone.0034144
Towler, A., White, D., & Kemp, R. I. (2014). Evaluating training methods for facial image
comparison: The face shape strategy does not work. Perception, 43(2-3), 214–218.
doi:10.1068/p7676
Troje, N.F. & Bulthoff, H.H. (1996). Face recognition under varying posers: the role of

251

texture and shape. Vision Research, 36, 1761 -1771.
Uttal, W, R., Baruch, T., Allen, L. (1995). Sequential image degradations in a recognition
task. Perception & Psychophysics, 57, 682-691.
Valentine, T. (1991). A unified account of the effects of distinctiveness, inversion, and
race in face recognition. Quarterly Journal of Experimental Psychology Section A, 43(2),
161-204.
Webster,M. & Mollen, J. (1995). Colour constancy influenced by contrast adaptation.
Nature. 373, 694-698.
Wells, G., Rydell, S. & Seelau, E. (1993). The selection of distractors for eyewitness
lineups. Journal of Applied Psychology. 78(5). 835-844.
Wells, G.L., Small, M., Penrod, S.D, Malpass, R.S., Fulero, S.M., Brimacombe, C.A.E. (1998).
Eyewitness identification procedures: recommendations for lineups and photspreads.
Law and Human Behaviour. 22, 603-607.
White, D., Burton, A. M., Kemp, R. I., & Jenkins, R. O. B. (2013). Crowd Effects in
Unfamiliar Face Matching, 777(November), 769–777.
White, D., Kemp, R. I., Jenkins, R., & Burton, A. M. (2014). Feedback training for facial
image comparison. Psychonomic Bulletin & Review, 21(1), 100–6. doi:10.3758/s13423013-0475-3
White, D., Kemp, R. I., Jenkins, R., Matheson, M., & Burton, A. M. (2014). Passport
officers’

errors

in

face-matching.

PloS

One,

9(8),

e103510.

doi:10.1371/journal.pone.0103510
White, D., Dunn, J. D., Schmid, A. C., & Kemp, R. I. (2015). Error rates in users of automatic
face recognition software. PLoS ONE, 10(10), 1–14. doi:10.1371/journal.pone.0139827

252

White, D., Phillips, P. J., Hahn, C. A., Hill, M., Toole, A. J. O., & White, D. (2015b).
Perceptual expertise in forensic facial image comparison, Proceedings of the Royal
Society, 282, 1–8.
Winograd, E. (1976). Recognition memory for faces following nine different judgments.
Bulletin of the Pscyhonomic Society, 8, 419-421.
Woodhead, M. M., Baddeley, a. D., & Simmonds, D. C. V. (1979). On Training People to
Recognize Faces. Ergonomics, 22(3), 333–343. doi:10.1080/00140137908924617
Yin, R. K. (1969). Looking at upside-down faces. Journal of Experimental Psychology, 81(1),
141–145.
Young, A.W., Hay, D.C., McWeeny, K.H., Flude, B.M. & Ellis, A.W. (1985). Matching
familiar and unfamiliar faces on internal and external features. Perception, 14, 737-746.

253