2012 Third International Conference on Emerging Security Technologies

Catching Even More Offenders with EvoFIT Facial Composites

Charlie D. Frowd, Melanie Pitchford, Faye Skelton, Anna Petkovic
University of Central Lancashire
[email protected], [email protected], [email protected], [email protected]

Colin Prosser and Brian Coates

Humberside police
[email protected] and [email protected]

Abstract construct a facial likeness. The resulting image is

Facial composites are an investigative tool used by police circulated within a police force or in the media for
to identify suspects of crime. Unfortunately, traditional identification—to generate further lines of enquiry and,
methods to construct the face have rather low success ideally, to locate the offender.
rates. We have been developing a new recognition-based Many techniques have been devised to construct faces
method called EvoFIT that requires eyewitnesses to select from memory. Initially, these were manual-type systems
whole faces from arrays of alternatives. Both published such as sketch, Identikit (in the US) and Photofit (UK).
laboratory research and existing police field-trials have Later, in the 1990s, software programs were created
found that EvoFIT produces images that are more including CD-FIT, E-FIT and PRO-fit; more recently
identifiable than images from traditional systems. In the there are US systems such as FACES and Identikit 2000.
current paper, we present an evaluation of a more recent All of these methods require witnesses to select individual
version of EvoFIT: in 2010, EvoFIT was deployed in 35 facial features (eyes, hair, nose, mouth, etc.) to create a
criminal investigations by Humberside police and these facial composition or ‘composite’ [1].
images directly led to identification of 21 suspects, Considerable research has established that these
equating to 60% success—quadruple the performance of ‘feature’ methods are ineffective when used to construct a
the previous system used within the force. The evaluation face from memory two or more days after a target face
also showed that identification of a suspect led to has been seen [1]. For example, in one of our projects [2]
conviction in 29% of investigations (6 out of 21). Overall, using police-like procedures and a 2-day delay, E-FIT and
a conviction occurred in 17% of cases involving use of an PRO-fit produced composites that were named by other
EvoFIT (6 out of 35). We also outline more recent people with an average (using the mean) of less than 5%
developments which indicate that an arrest is now likely correct. This finding has been replicated (e.g. [3][20])
in three out of every four cases in which EvoFIT is used, and found to apply to other feature systems [5][6].
and a conviction rate of one in five. It is for this reason that alternative methods have been
developed to construct facial likenesses [7][8][9]. These
are based on selecting complete faces, a task that should
1. Introduction be better aligned to natural (holistic) face recognition
[12], and potentially easier for a witness to do than
Sometimes, physical evidence is available in police selecting individual features. We have been developing
investigations to allow suspects to be identified: CCTV EvoFIT ([1][3][6][9][10][11][13][14][18][19]), in which,
footage, fingerprints, DNA, footprints, etc. At other witnesses are shown screens of complete faces and select
times, however, the evidence is less tangible, being those that resemble the offender. Selected faces are
contained in the memory of an eyewitness. In this latter combined together (similar to genetic ‘breeding’) to
situation, recovering an accurate account of what produce more faces for selection and, when repeated a
happened and the people involved are important for few times, a composite is ‘evolved’. While the approach
solving the case. For serious crime—rape, burglary and has intuitive appeal, about 10 years of intensive research
murder—investigating officers normally consider were required to develop a system that could produce
construction of a facial composite in order to identify a more identifiable composites than other systems [1].
suspect. Here, witnesses and victims who have clearly The technical details of EvoFIT have been described
seen the offender’s face describe the appearance of the in many research papers and system reviews. See Frowd
face and work with a trained composite officer to et al. [9] for a thorough technical review of its
development, computer simulations and initial testing

978-0-7695-4791-6/12 $26.00 © 2012 IEEE 20

DOI 10.1109/EST.2012.26
with users, and [18] and [19] for more recent laboratory for identification by another personn [4]. It also contained
evaluations. We provide a brief technical account here. ‘holistic’ tools to manipulate the evolved face: to adjust
At the heart of EvoFIT is a face generrator built using perceived age, masculinity, heallth and other overall
Principal Components Analysis (PCA A), to capture properties of the face; further scalles are available to add
variation in shape and texture for a databbase of about 70 various types of facial hair, and lighten and/or darken
facial images (of a given race, gender and age range). eyes, brows, etc. Using the same construction procedures,
Initial faces presented to witnesses are gennerated from this including a 48-hour delay to interrview, composites from
computer model with random PCA param meters, in sets of the PRO-fit feature system were naamed at 5% [3]. At this
18. Figure 1 is an example set producedd from the white time, then, identification was aboutt five times higher from
male database covering ages of 30 to 40 yyears. Witnesses EvoFIT than from a leading featurre system. Even better
select from screens such as these, and the parameters performance is now possible (see below
b and section 3.1).
underlying these selections are combinned together to From 2007 to 2009, field triaals were carried out in
produce more faces for witnesses to seleect. Technically Lancashire and Derbyshire police forces [13]. In spite of
speaking, this ‘breeding’ process invvolves uniform the difficulties involved in running field studies, discussed
crossover and selection with parameteers occasionally in detail in [13], methods were esstablished for assessing
mutated (to increase variability within thhe population of performance. Police officers werre trained within these
faces, improving performance). Witnessses also select a forces and used EvoFIT for a nominaln period of six
single face that represents the best overrall likeness and months. EvoFIT composites directtly lead to identification
this preferential item is given greater weigghting: twice the and arrest of a suspect in 20% of cases—see
c Figure 2 for
number of breeding opportunities inn the Genetic an example case. In contrast, the previous E-FIT system
Algorithm; it is also transferred unalterred to the next did not provide any identifications1.
generation as part of an ‘elitist’ approach. The process of
selection and breeding is repeated, to ‘evolve’ a face.

Figure 2. An EvoFIT (left) construccted by a victim during

the Lancashire police field trials. The
T image was named
as Ross Gleave (right) and led to hiis arrest and conviction
of attempted rape of a child underr 13 years.

From summer 2009 to spring 2010,

2 further field trials
were run by Devon and Cornwall police, and police in
Figure 1. An example EvoFIT screenshot. Witnesses and Romania. These evaluations were similar in nature to the
victims are shown screens such as thesee from which to above but used a simplified proceedure in the face-recall
make selections, choosing those faces wh hich look overall interview with witnesses. Previoussly, witnesses described
like the face of the offender. Selected faces are the appearance of the offender’ss face, known as free
combined together by the system for wittnesses to make recall, and were then prompted d to remember more
further selections. In the latest version of the software information about it using a standarrd question-and-answer
[19], to promote optimal identification off the composite, format, cued recall. Howeverr, based on research
the focus is first on the central part of thee face (as shown suggesting potential problems witth this latter technique
here); later, witnesses add the external parts (hair, ear
and neck) to produce a complete face. 1
For the 2007-10 field trials [13][14], EvoFIT was used 126 times and
identified 32 suspects; in a comparable audiit of E-FIT, no identifications
In 2007, volunteers produced com mposites using resulted from 20 composites. This increase in identification from E-FIT
EvoFIT with a mean naming of 25% coorrect [3]. This to EvoFIT is statistically reliable (at the standard
s probability level of
software contained techniques to help wittnesses focus on 0.05) using a Fisher Exact Probability Tesst (p = .027). (N.B. A Chi-
Square test is not appropriate here as one of
o the expected frequencies is
the central part of the face—the region tthat is important less than five.)

21
[20], these later field trials did not invoolve cued recall. 2. Humberside police
Identification was found to be markedlyy better, with a Humberside police is one of 43
3 forces in England and
suspect identified in 38% of cases in Rommania and 40% in Wales. They are based in No orth-East England and
Devon and Cornwall. See Figure 3 for an example case. comprise areas in the counties of Lincolnshire and
Yorkshire, with major towns off Kingston upon Hull,
Grimsby and Scunthorpe. To policce this region, the force
has four Basic Command Units, or Divisions.

2.1. Humberside’s previous co

omposite system

Humberside police used the E-FIT feature system in

investigations from 2007 to 2009 9. During this period,
seven police officers worked with h witnesses and victims
to produce 56 composites for cases c of murder, rape,
sexual assault and burglary. Thesee 56 E-FITs led to eight
identifications in total—a name put forward that, in
conjunction with other evidence, produced eight suspects.
The level of suspect identificationn was therefore 14% (8
Figure 3. An EvoFIT constructed of a violeent robber (who out of 56). In 2009, no identificatiions were produced and
also used a weapon) and a photograph (right) of the so the force considered alternativee composite systems to
person convicted of this and a second sim
milar offence. make better use of police time, effo
ort and money.

The field trials also made use of a neew technique for 2.2. EvoFIT evaluation and fieeld trial
police to make public appeals on TV, wanted-persons
web-pages and online newspapers. The technique In 2010, Humberside police conducted an initial
presents a composite as a series of proggressively more- evaluation of EvoFIT. Very-good d performance resulted
exaggerated caricatures, similar to the w way that an artist and EvoFIT was rolled out acrosss the force; it was also
caricatures a face. When observing suuch a sequence, deployed about twice as often annually
a than was the
identification increases substantially [15]. It is a standard previous system. Identification off suspects was 60%, as
animation format and is produced by a com mposite operator described below.
as a secondary exhibit for enhancement ((and a statement
is produced to this effect). An exampple can be seen 2.2.1. Initial evaluation. It is normal
n for UK police
online at http://www.uclan.ac.uk/animateddcomposite. forces to evaluate products beforee purchase, and facial-
Research continued to seek ways to im mprove EvoFIT, composite software normally fo ollows this procedure.
and in 2010, we developed an even more eeffective method Two existing composite operattors were trained on
[19]. This involved witnesses first cconstructing the EvoFIT in January 2010 and evaluaated the system for four
internal facial features, the region incluuding the eyes, months. During this time, six composites were
brows, nose and mouth, and which aree important for constructed for burglaries, and one
o each for a sexual
successful recognition of the face [4]; havving created this touching case and to locate a missing
m person. Five
region, witnesses then select hair, eears and neck. identifications (arrests) were made from names put
Laboratory tests indicated that EvoFITs pproduced in this forward from these images.
way were twice as identifiable, at 45% ccorrect, than the
previous method. An updated version off the system was 2.2.2. Operationalizing EvoFIT. Based
B on the success of
given to police users in 2010 to pilot teest and formally the initial evaluation, eight furtheer police officers were
audit. trained, to give appropriate cover in the four Divisions.
The current paper describes a more recent audit of Five of these officers had already been trained on E-FIT
EvoFIT’s performance, specifically for a 112-month period and so were given the same traaining as the first two
in Humberside police starting Januaryy 2010. It also officers. This was a 2-day coursee covering interviewing
indicates performance using the above ‘internals-first’ methods for face construction with EvoFIT (e.g. including
method of construction. In the final part of the paper, an free but not cued recall), in nstruction on EvoFIT
overview is provided of more recent enhhancements and procedures and holistic- and featu ure-manipulation tools,
system characteristics. with in-class practice. Another trraining course was run
for three further officers withou ut previous composite-
operator training. These trainees attended
a a 5 day course
that also gave instruction on histo ory of composites, UK

22
national police guidelines for compossite production, FIT (8 out of 56) used in force duriing the previous 3 years
exhibiting composites as evidence in coourt and use of [X2(1) = 8.9, p = .003]; the Odds Ratio (a standard
appropriate artwork techniques—for addding extra detail measure of effect size) for thiis comparison is 4.2,
such as moles, scars, unusual facial hair annd tattoos. meaning that an EvoFIT was 4.2 tim mes more likely to lead
to an identification than was an E-FFIT.
2.2.3. Audit of performance. EvoFIT w was deployed 35 Six identifications led to conv viction—three were for
times during the 12-month evaluation. F Figure 4 presents distraction burglary, and one each for aggravated
the distribution of offences for composiites constructed. burglary, burglary in a dwelling an nd assault. The result is
Burglary in general represented about twoo-thirds (63%) of that 29% of identifications (6 out of
o 21) led to conviction.
cases; the category of ‘Other burglary’ inncludes burglary So, for all 35 EvoFIT compossites produced in the
with grievous bodily harm (GBH, one caase), aggravated evaluation, 17% of them (6 out of 35)3 led to conviction.
(involving use of a weapon, one case) and (attempted) From the available data2, the shortest interval from
distraction burglary (two cases). Robbbery and assault onset-of-crime to construction-of-E EvoFIT was 1 day, and
(involving Actual Bodily Harm) equate too 11% each and this occurred four times (13%). The longest was for a
the remaining cases (14%, ‘Other’) are ffor a Section 18 witness interviewed after two mon nths (who produced two
offence (causing GBH with intent), attemppted rape, sexual composites)—such an interval occurs o occasionally in
touching, and for locating a missing personn. criminal investigations; for the following analysis, to
avoid skewing the data, this outly ying case is removed3.
For the remaining EvoFITs (Totall, N = 30), the average
interval to construction was 3.8 day ys (Standard Deviation,
SD = 2.4 days). The average was very similar for the 21
cases involving identification of a suspect
s (Mean, M = 3.8
days, SD = 2.5 days) and for th he 6 cases that led to
conviction (M = 3.7 days, SD = 1.8 1 days). Intervals are
not statistically different for caases with or without
7], nor for identification
identification [t(28) = 0.03, p = .97
with and without conviction [t(19) = 0.2, p = .87].
Overall, 50% of EvoFITs were con nstructed within two to
three days of the crime, and 93% of them were
Figure 4. Distribution of 35 crimes wheree an EvoFIT was constructed within a week.
constructed in the trial. Data are arran nged by type of
offence. See text for definitions and moree details. 2.2.4. Case studies. In this sectiion, we consider cases
which provide the strongest evid dence that EvoFIT had
An audit was carried out by the force 12 months after correctly identified the perpetrator of a crime—that is, for
the evaluation period. Out of 35 EvoFIITs constructed, investigations leading to convictio on. For example, one
several composites were not used for ideentification (e.g. case relates to a serious assault onn the owner of a hotel.
suspect subsequently identified by other m means). In total, Here, one of the two suspects was w identified through
21 suspects were directly identified from m the EvoFITs. publication of the EvoFIT, and thee enquiry revealed that
Identification of suspects by type of crim me is shown in the victim’s business partner had orrganised the attack.
Table 1. Assault (with ABH) resultedd in the highest Figure 5 illustrates composites constructed from three
successful identification (4 out of 4); the ‘‘Other’ category separate investigations. The EvoF FIT is shown on the left
contained a single success, identifying a suspect in an and the person convicted in that case is shown on the
attempted-rape case. For the largest grouup of composites right. The EvoFIT in the first ro ow was produced by a
constructed, for all types of burglaryy, the average female victim of distraction-burglaary. In the second row,
identification was 59% (13 out of 22). the EvoFIT was constructed by b a male victim of
aggravated-burglary. In this crim me, the victim returned
Table 1. Identification of EvoFIT composittes (expressed in home to find a male burglar in his h house. The burglar
percentage of total per category) as a funnction of type of made attempts to assault him with h an axe before making
crime. The categories are the same as thoose in Figure 4. off. Both EvoFITs in the third and fourth row were
Distraction Other produced by a female victim of disttraction burglary.
Burglary Robbery ssault
As Other
Burglary Burglary
63% 50% 75% 75% 1
100% 20%
2
At the time of writing, we were awaiting g confirmation of dates from
Overall identification was 60% (21 ouut of 35). Using three cases that did not lead to identification of a suspect, and so our
Chi-square statistics, this level of pperformance is analyses for construction interval relate to 32 out of 35 investigations.
3
significantly higher than the combined peerformance of E- We use the established method [17] of removal
r based on 2 standard
deviations from the mean (for non-identificaation cases).

23
 3. Discussion
In the absence of suitable leaads, it can be of vital
importance for effective policing to obtain an identifiable
image of an offender’s face from m eyewitness memory.
Systems for constructing compositees have come and gone,
but none have been capable of doing the one thing for
which they were intended: to producep an image that
another person could correectly identify (e.g.
[2][3][5][14]). We have been atteempting to achieve this
objective since 1998. The resultt is a software system
which synthesises arrays of faces fa from computer-
generated models. Witnesses repeaatedly select faces from
such arrays, with ‘breeding’ of sellected items, to allow a
composite to be ‘evolved’. Afterwards,
A there are
‘holistic’ tools available to improvee the overall likeness of
the face and to manipulate individu ual features on request.
In addition, the initial focus is on n the important central
part of the face, the region that t is important for
recognition of the composite by ano other person later.
In 2007, we developed an Evo oFIT system that could
produce composites that other people p named at 25%
correct [3]. Field trials of this system
s [13] revealed a
similar figure for identifications leeading to the arrest of a
suspect; Figure 2 illustrates a high profile case that led to
conviction at the time. The processs used with EvoFIT has
been improving [1], and a further field
f trial [14] found an
arrest rate of about 40% wheen police interviewed
witnesses using the simplified face-recall procedure
outlined earlier.
In the current paper, we descrribe an evaluation of a
more-recent version of EvoFIT. Over O a 12-month period
in 2010, 35 composites were consstructed by Humberside
police with witnesses and victims of o mostly serious crime,
about two-thirds of which were fo or burglaries. Twenty-
one suspects were identified, or 60% of all EvoFITs
constructed in the trial. This perrformance is over four
times higher than composites from m the previous system
(E-FIT) used by the force. It was also
a found that six cases
led to conviction, meaning that 29% % of identifications led
to conviction, or 17% of all EvoFIITs. These latter figures
reflect the situation that many cases do not have sufficient
corroborative and physical eviden nce (DNA, fingerprints,
fibres, etc.) to ensure a reliable coonviction, and therefore
do not proceed to court.

3.1. Further research and deveelopment

Further research has resultted in two notable

developments for EvoFIT. The first is an improvement to
the initial interview used immeediately prior to face
construction. It is normal for an operator to chat
Figure 5. EvoFIT composites constructeed (left) in the informally with witnesses, to builld rapport, and then to
Humberside police evaluation and the relevant person ask them to think back to the crimee and visualize the face,
who was convicted (right). See text for m
more details. context reinstatement; next, if wiitnesses are able, they
provide a free-recall description n of the face. As

24
mentioned above, operators may also prompt witnesses to and female of different ages and races—White, Black,
attempt further recall on individual features, cued recall— Asian, Eastern European, Chinese, Hispanic and various
although this technique seems to be detrimental to face mixed-race combinations. A common procedure for face
construction [14]. However, face recall in general may construction is used with each database using a holistic
encourage witnesses to focus on individual features cognitive interview and holistic and feature tools; finished
during face construction, which is unlikely to be an images can be published in the media as animated and/or
optimal strategy since an identifiable image requires perceptual-stretch composites.
accurate features along with accurate placement of
features on the face—to give an overall (holistic)
impression. The interview was enhanced to overcome
this issue: after face recall, witnesses now focus on the 4. Summary
overall properties or character of the face. This ‘holistic’
cognitive interview (H-CI) is straightforward to learn and About 15 years of intensive laboratory research and
administer, but allows a witness to produce a substantially police fieldwork has been dedicated to developing
more identifiable EvoFIT [18]. EvoFIT. The designers have sought, using procedures as
The second improvement is a new method for used in policework, to produce identifiable composites.
publishing composites in the media as part of a public The EvoFIT team has provided police training and after-
appeal for information. The reader may recollect that sales (technical) support since 2007.
animated composites are useful for TV and internet Here, we described results of the latest field trial with
outlets [15], and were used in the current evaluation. The Humberside police in 2010. The force constructed 35
new technique is an additional secondary exhibit that can EvoFIT composites with witnesses in a range of serious
be used for the newspapers. It is called ‘perceptual- crimes and identification of a suspect (an arrest) directly
stretch’ and simply asks observers to look at the resulted from these images in 60% of cases; 29% of these
composite from the (left- or right-hand) side. When this arrests led to conviction. Put another way, for all
is done, the face appears to be longer-and-thinner than composites produced in the evaluation, EvoFIT led to
normal—in other words, it is stretched vertically. Our identification and then conviction once in every six cases.
cognitive system appears to deal with this situation by Further developments have been made to improve system
normalising the face and, in doing so, is less sensitive to performance further and field trials are under way to
some of the errors in the face, improving recognition. assess the current level of suspect identification and
In a recent laboratory test, the current version of conviction. The project is an excellent example of the
EvoFIT was evaluated in conjunction with the H-CI and benefit to be gained by careful laboratory research
perceptual-stretch technique for identification. Forty coupled with collaborative formal field trials to develop
volunteers looked at a short video clip of an unknown an effective tool for policing.
face, and 24 hours later constructed an EvoFIT composite
of it either after the normal face-recall interview, or
following an H-CI. As found before [3], EvoFITs were 5. References
correctly named at 25% using the older techniques, but
naming increased to 56% using H-CI, and then again to [1] C.D. Frowd, V. Bruce, and P.J.B. Hancock, “Evolving facial
74% using both H-CI and perceptual stretch. composite systems”, Forensic Update, 98, 2009, pp. 25-32.

3.2. Current version of EvoFIT [2] C.D. Frowd, D. Carson, H. Ness, … and P.J.B. Hancock,
“Contemporary Composite Techniques: the impact of a
forensically-relevant target delay”, Legal and Criminological
Laboratory research and police field trials have Psychology, 10, 2005, pp. 63-81.
enabled creation of a system capable of producing very
identifiable composites. Further field trials are currently [3] C.D. Frowd, M. Pitchford, V. Bruce, S. Jackson, G. Hepton,
in progress to measure the extent of the above and newest M. Greenall, A. McIntyre, and P.J.B. Hancock, “The psychology
developments, but based on indications so far (previous of face construction: giving evolution a helping hand”, Applied
paragraph), an arrest should now result in three out of Cognitive Psychology. 2010, DOI: 10.1002/acp.1662.
every four cases where EvoFITs are produced; also, since
29% of these cases successfully proceed further (Section [4] H.D. Ellis, J.M. Shepherd, and G.M. Davies, “Identification
of familiar and unfamiliar faces from internal and external
2.2.3), this means that one in five EvoFITs should result
features: some implications for theories of face recognition”,
in conviction (or 21%, calculated as 29% convictions for Perception, 8, 1979, pp. 431-439.
74% identifications).
It would appear then that EvoFIT makes excellent use [5] C.E. Koehn, and R.P. Fisher, “Constructing facial
of police time and resources. There are now 60 individual composites with the Mac-a-Mug Pro system”, Psychology,
face databases, for production of different offenders: male Crime and Law, 3, 1997, pp. 215-224.

25
[6] C.D. Frowd, D. McQuiston-Surrett, S. Anandaciva, C.E. [13] C.D. Frowd, P.J.B. Hancock, V. Bruce, A. McIntyre, M.
Ireland & P.J.B. Hancock, “An evaluation of US systems for Pitchford, R. Atkins, et al. (2010), “Giving crime the 'evo':
facial composite production.” Ergonomics, 50, 2007, pp. 1987- catching criminals using EvoFIT facial composites”. In G.
1998. Howells, et al. (Eds.) 2010 IEEE International Conference on
Emerging Security Technologies, 2010, pp. 36-43.
[7] S.J. Gibson., C.J. Solomon & A. Pallares-Bejarano,
“Synthesis of photographic quality facial composites using [14] C.D. Frowd, P.J.B. Hancock, V. Bruce, F. Skelton, C.
evolutionary algorithms.” In R. Harvey and J.A. Bangham Atherton, … G. Sendrea, “Catching more offenders with
(Eds.) Proceedings of the British Machine Vision Conference, EvoFIT facial composites: lab research and police field trials”,
2003, pp. 221-230. Global Journal of Human Social Science, 11, 2011, pp. 46-58.

[8] C.G. Tredoux, D.T. Nunez, O. Oxtoby, and B. Prag, “An [15] C.D. Frowd, V. Bruce, D. Ross, A. McIntyre, & P.J.B.
evaluation of ID: an eigenface based construction system”, Hancock “An application of caricature: how to improve the
South African Computer Journal, 37, 2006, pp. 1-9. recognition of facial composites”, Visual Cognition, 15, 2007,
pp. 1-31.
[9] C.D. Frowd, P.J.B. Hancock, and D. Carson, “EvoFIT: A
holistic, evolutionary facial imaging technique for creating [16] C.D. Frowd, D. Carson, H. Ness, J. Richardson, L.
composites”, ACM Transactions on Applied Psychology (TAP), Morrison, S. McLanaghan, and P.J.B. Hancock, “A forensically
1, 2004, pp. 1-21. valid comparison of facial composite systems”, Psychology,
Crime and Law, 11, 2005, pp. 33-52.
[10] C.D. Frowd, V. Bruce, C. Gannon, M. Robinson, C.
Tredoux, J. Park., A. McIntyre, and P.J.B. Hancock, “Evolving [17] M.V. Selst, and P. Jolicoeur, “A solution to the effect of
the face of a criminal: how to search a face space more sample size on outlier elimination”, Quarterly Journal of
effectively”. In A. Stoica, T. Arslan, D.Howard, T. Kim and A. Experimental Psychology Section A, 47, 1994, pp. 631-650.
El-Rayis (Eds.) 2007 ECSIS Symposium on Bio-inspired,
Learning, and Intelligent Systems for Security, 2007, pp. 3-10. [18] C.D. Frowd, L. Nelson, F.C. Skelton, R. Noyce, P. Heard,
J. Henry, … and P.J.B. Hancock, “Interviewing techniques for
[11] C.D. Frowd, V. Bruce, C. Gannon, M. Robinson, C. Darwinian facial composite systems”, Applied Cognitive
Tredoux, J. Park., A. McIntyre, and P.J.B. Hancock, “Effecting Psychology, 2012, DOI: 10.1002/acp.2829.
an improvement to the fitness function. How to evolve a more
identifiable face.” Evolving the face of a criminal: how to search [19] C.D. Frowd, F. Skelton, C. Atherton, C., M. Pitchford, …
a face space more effectively. In A. Stoica, T. Arslan, and P.J.B. Hancock, “Recovering faces from memory: the
D.Howard, T. Kim and A. El-Rayis (Eds.) 2007 ECSIS distracting influence of external facial features”, Journal of
Symposium on Bio-inspired, Learning, and Intelligent Systems Experimental Psychology: Applied, 2012, DOI:
for Security, 2007, pp. 3-10. NJ: CPS. 10.1037/a0027393.

[12] J.W. Tanaka, and M.J. Farah, “Parts and wholes in face [20] C.D. Frowd, and S. Fields, “Verbal overshadowing
recognition”. Quarterly Journal of Experimental Psychology: interference with facial composite production. Psychology,
Human Experimental Psychology, 46A, 1993, pp. 225-245. Crime and Law, 2010, DOI: 10.1080/10683161003623264.

26