Content uploaded by Gary L. Wells
Author content
All content in this area was uploaded by Gary L. Wells on Oct 31, 2017
Content may be subject to copyright.
UNCORRECTED PROOF
Facial Composite Production by
Eyewitnesses
Gary L. Wells and Lisa E. Hasel
Iowa State University
ABSTRACT—The creation of facial images by eyewitnesses
using composite-production systems can be important for
the investigation of crimes when the identity of the per-
petrator is at issue. Despite technological advances,
research indicates that composite-production systems
produce poor likenesses of intended faces, even familiar
faces. Furthermore, producing a composite appears to
harm later recognition performance. Although morphing
composites from multiple witnesses helps, likeness is still
limited. The problem might stem from a mismatch between
how faces are represented in memory (holistically) and
how composite systems attempt to retrieve the memories (at
the feature level). New methods of face recall involving
judgments of whole faces hold greater promise.
KEYWORDS—eyewitnesses; composites; face recall
Kirk Bloodsworth, a U.S. military veteran living in Maryland,
had never been in trouble with the law. Nevertheless, he was
convicted of the 1984 rape and murder of a 9-year-old girl and
was sentenced to die in Maryland’s gas chamber. Bloodsworth
became a suspect in the murder because an anonymous person
called police to say Bloodsworth looked like a composite face
that police had released to the media. Lacking a compelling alibi
for the time of the crime, police placed his photo in a photo
lineup and eyewitnesses identified him. After 9 years in prison,
DNA evidence vindicated Bloodsworth and also implicated the
actual murderer, Kimberly Ruffner (Junkin, 2004). Ruffner did
not look much like the composite, but Bloodsworth did, and
Bloodsworth was the only one in the photo lineup with hair that
matched the composite.
The Bloodsworth case illustrates some key points in this article.
First, laboratory research shows that a face composite by an eye-
witness is generally a poor representation of the original face.
Hence, a composite has the potential to lead crime investigators
away from the real perpetrator and toward an innocent person. Also,
a composite can bias the eyewitness away from identifying the
original face and toward a face that resembles the composite. We
review research on face composites, explore the question of why
people are not better at building them, examine a new approach to
face recall, and underscore the need for psychological science to
help address this important problem in the justice system.
COMPOSITE-PRODUCTION SYSTEMS
The objective of many criminal investigations is to establish the
identity of the perpetrator. When there is a suspect, eyewitnesses
can help establish identity by viewing a lineup that contains that
suspect. When there is no suspect in the case, however, inves-
tigators often rely on eyewitnesses to help produce a likeness of
the perpetrator’s face. The first method for having eyewitnesses
produce a face from memory was the sketch artist. Today, how-
ever, U.S. law enforcement agencies use mechanized systems
and are over twice as likely to use computerized versions than
noncomputerized ones (McQuiston-Surrett, Topp, & Malpass,
in press).
Studies comparing sketch artists to mechanized systems are
rare, perhaps because sketch artists vary widely in their skills
and, hence, it would take a large sample of sketch artists ran-
domly sampled from an ill-defined population to make conclu-
sions. Accordingly, we restrict our review and discussion to
composite-production systems such as the Identi-Kit, Photo-Fit,
E-Fit, Mac-a-Mug, and FACES. The first two are early, non-
computerized collections of facial features (e.g., noses, eyes,
mouths, head shapes, hair styles) that can be superimposed to
create a face. The latter three are examples of modern, compu-
terized versions of the same idea but include more possible facial
features and more realistic visual results. FACES, for example,
includes 361 hair selections, 63 head shapes, 42 forehead lines,
410 eyebrows, 514 eyes, 593 noses, 561 lips, 416 jaw shapes, 145
moustaches, 152 beards, 33 goatees, 127 eyeglasses, 70 eye
lines, 147 smile lines, 50 mouth lines, and 40 chin lines. Figure 1
CDIR 465 BDispatch: 31.1.07 Journal: CDIR CE: Blackwell
Journal Name Manuscript No. Author Received: No. of pages: 5 Saravan A/Anand Kumar
Address correspondence to Gary L. Wells, Psychology Department,
West 112 Lago, Iowa State University, Ames, IA 50021; e-mail:
glwells@iastate.edu.
CURRENT DIRECTIONS IN PSYCHOLOGICAL SCIENCE
6Volume 16—Number 1Copyright r2007 Association for Psychological Science
CDIR 465
(BWUS CDIR 465.PDF 31-Jan-07 15:59 907858 Bytes 5 PAGES n operator=M.V.Anantha)
UNCORRECTED PROOF
shows a potential composite-building sequence using the FACES
program. Each panel shows one or more features added in a se-
quence of steps. The eyewitness can begin with any feature and
add features in any order. The upper left panel is hair only; the
upper middle panel adds eyes; the upper right panel adds head
shape and eyebrows (two steps); the lower left panel adds a nose;
the lower middle panel adds lips and jaw shape (two steps); and
the lower right panel adds forehead lines, eye lines, smile lines,
mouth lines, and chin lines (five steps). A witness can go back
and change any feature (e.g., eyes) by selecting a different one
and the new feature replaces the old one.
The evolution of face-composite systems has generally in-
volved increasing the number of available features, increasing
the realism of the composite produced, and increasing the user-
friendliness of the system. Recently, a new generation of face-
recall systems has emerged that moves away from the feature-
selection method of previous composite systems and instead
uses whole faces. In one sense, these new systems, discussed
later in this article, are not really composite systems at all, but
they might hold greater promise for face recall than that of their
predecessors.
HOW GOOD ARE COMPOSITES?
Research on the ability of people to use composite systems to
produce likenesses of intended faces has a 30-year history, be-
ginning with researchers in the United Kingdom. The results
were disappointing from the outset (e.g., see Davies, Ellis, &
Shepherd, 1978), and modern, computerized versions have not
fared much better (Davies & Valentine, 2006). Many methods of
assessing the accuracy of face composites have been developed,
including matching tasks (‘‘Which of these faces was this com-
posite intended to depict?’’), naming tasks (e.g., ‘‘What famous
person is this?’’), and similarity-rating tasks (‘‘How similar is this
composite to this actual face?’’). One of the problems is the
difficulty of comparing these measures across different studies,
because each measure tends to depend on context. Matching
tasks, for example, are sensitive to the similarity among the al-
ternatives presented and the set size of the alternatives. Even if
chance were defined as 25% in a four-alternative measure, a
given composite might show a 60% match rate for four alterna-
tive faces if the faces were very different from one another or a
30% match rate if the alternative faces were more similar to each
other.
If it is difficult to compare measures across studies, it is even
more difficult to describe some ‘‘absolute’’ level of performance.
So, why have researchers tended to conclude that composite-
production systems typically produce poor likenesses of the
intended targets? Poor relative to what? Clearly, this is a judg-
ment call, but we agree with this overall conclusion based on two
general observations. First, studies in which individuals at-
tempted to create composites of famous faces show that very few
people recognize the composite face even though they know the
face of the famous person very well. Frowd et al. (2005), for
instance, found that participants correctly named only 2.8% (22
of 800 possible) of a group of 50 famous faces based on
Fig. 1. Progressive sequence of building a composite face using the FACES 3.0 (IQ Biometrics) program. The eyewitness begins with any feature
and adds features in any order. In this case, hair (upper left) is followed by eyes (upper middle); head shape and eyebrows (upper right); nose
(lower left); lips and jaw shape (lower middle); and forehead lines, eye lines, smile lines, mouth lines, and chin lines (lower right).
Volume 16—Number 1 7
Gary L. Wells and Lisa E. Hasel
CDIR 465
(BWUS CDIR 465.PDF 31-Jan-07 15:59 907858 Bytes 5 PAGES n operator=M.V.Anantha)
UNCORRECTED PROOF
composites of those faces that had been created by other indi-
viduals. Similarly, people seem largely unable to recognize
people they personally know from composites of those individ-
uals. Kovera, Penrod, Pappas, and Thill, (1997), for example,
found that people were unable to discriminate between com-
posites of their classmates they knew and composites of students
from other schools whom they did not know. In general, studies of
composite production reveal that participants create poor like-
nesses of intended faces, regardless of whether the composite
builder knows the target face or has had only a brief encounter
with that person, and regardless of the type of task (naming,
matching, or similarity rating) used to assess the likeness (see
Davies & Valentine, 2006, for an excellent review).
Recent research has shown that there can be benefits to
morphing (averaging at the pixel level) multiple composite faces
of the same individual created by different eyewitnesses (Bruce,
Ness, Hancock, Newman, & Rarity, 2002; Hasel & Wells, 2007).
Sampling across 16 different strategically sampled target faces,
Hasel and Wells (2007) found that morphing composites from
four independent participants, each of whom had viewed the
same target face, produced a new image that was a better like-
ness of the target than was the average individual composite.
Judges rated the similarity of each composite to the target face
plus three distractor faces. Using the individual composites, the
target received the highest rating 35% of the time (chance 5
25%). Using morphed composites, however, the target received
the highest rating 48% of the time. The superiority of the morph
is evident in Figure 2, which depicts an example from the Hasel
and Wells experiment. The figure shows one of the 16 target
faces, four composites prepared by separate participants, meant
to depict that target, and the morphed result from the four
composites. The example in Figure 2 is one of the better results
across the 16 target faces, but it illustrates the general pattern we
observed: Morphing the composites produces a better likeness of
the target than any one of the individual composites. Participants
who built composites also rated how good they thought their own
individual composites were. Those receiving the highest ratings
were not more similar to the target faces than were the average
individual composites, indicating that the composite builders
were poor judges of how good or bad their own composites were.
Target face
Morph of the four composites
Composite 1
Composite 3
Composite 4
Composite 2
Fig. 2. A target face, four composites made by separate individuals, and the
morph of the four composites. From the Hasel & Wells (2007) study.
8Volume 16—Number 1
Facial Composite Production by Eyewitnesses
CDIR 465
(BWUS CDIR 465.PDF 31-Jan-07 15:59 907858 Bytes 5 PAGES n operator=M.V.Anantha)
UNCORRECTED PROOF
WHY ARE COMPOSITES POOR REPRESENTATIONS OF
THE INTENDED FACE?
No composite system could have enough facial-feature choices
to represent the physiognomic variability of the human face. But,
is the absence of enough features the problem with composite
systems? Not likely. Wells and Hryciw (1984), for example,
showed participants target faces that were created using the
Identi-Kit and the participants also used the Identi-Kit for
building their composites. Hence, it would have been possible to
perfectly reconstruct the target face because every one of its
features was available for selection from the composite kit.
Nevertheless, resemblance scores from judges averaged less
than 2.0 on a 7-point scale (1 5does not resemble, 7 5closely
resembles). Furthermore, despite incredible advances in the
technology of composite systems, including large increases in
the number of facial features available, enhanced realism such
that composites are almost indistinguishable from actual faces,
and varied methods of approaching the task of building the
composite, these technological advances have resulted in no
consistent improvement in composite likenesses to the original
faces (Davies & Valentine, 2006).
The dominant view in the research literature for over two
decades has been that there is a mismatch between the task
demands that are somewhat inherent in all composite systems
and the way that faces are usually perceived and remembered.
Numerous lines of evidence converge on the view that faces are
generally processed, stored, and retrieved at a holistic level
rather than at the level of individual facial features (see Tanaka
& Farah, 2003). There are various views of what is meant by
holistic, but the general idea is that the psychological process is
not merely a processing of individual facial features (e.g., eyes,
nose, mouth). Instead, faces might be represented in terms of
their multidimensional similarity to other known faces (e.g.,
Valentine, 1995) or in a coordinate spatial relations system that
includes distances between features, relative sizes of features,
and so on that cannot be separated from the features themselves
(Cooper & Wojan, 2000). As a result of holistic processing, recall
of individual facial features is rather poor; and yet face-com-
posite systems require the individual to recall individual facial
features.
This is not to say that whole faces cannot be processed at the
feature level if the task demands it. Wells and Hryciw (1984) had
participants evaluate faces for 10 personality traits (e.g., how
honest is this person?), a task that is likely to be performed based
on whole-face processing, or evaluate faces for 10 physical-
feature characteristics (e.g., how large is this person’s nose?), a
task that forces processing of isolated facial features. Later,
participants either attempted to recognize the faces they had
evaluated from six-person photo lineups or use the Identi-Kit to
build the faces from memory. Those who rated the faces for
personality traits performed much better on the lineup recog-
nition tasks than did those who rated the individual features, but
the opposite result occurred for the Identi-Kit task. This result is
consistent with the idea that holistic processing (encouraged by
personality-trait encoding) helps later recognition of the whole
face but harms composite task performance whereas feature-
based processing harms later recognition of the whole face and
helps composite task performance.
A NEW SYSTEM APPROACH TO FACE RECALL
The view that faces are normally processed and stored in some
type of holistic fashion rather than stored as constituent parts has
led researchers to develop whole-face methods for face recall.
These systems are still under development. They begin by
generating a random set of faces and the witness selects the face
most similar to memory for the target face. This face becomes the
‘‘parent’’ face that yields a new set of faces that are mutations of
that face, using any of a number of possible algorithms, and the
witness again makes a choice. This process is repeated until the
witness cannot choose because all the faces resemble the target
face equally well. Systems of this sort have been developed by
Hancock (2000) and by Gibson, Pallares-Bejarano, & Solomon
(2003). Comparison of these systems to traditional composite
systems has thus far been very limited. The few tests thus far
have not shown these particular versions of whole-face sys-
tems to be superior to traditional composite systems (Davies &
Valentine, 2006). Nevertheless, these whole-face systems rep-
resent a radically different approach to producing composite
faces and seem to hold the best prospects for a breakthrough
because they rely on the idea, derived from basic research on
face processing, that faces are processed holistically.
FINAL COMMENTS
The difficulty that people have with constructing face compos-
ites is not the result of weak memory for faces per se. People
produce poor composite likenesses even for faces that they know
very well and can easily recognize (Frowd, et. al., 2005). Instead,
it appears that human face processing is designed more for face
recognition, which is facilitated by holistic representations, than
it is for face recall, which requires individual feature repre-
sentations. Within a few days of birth, babies can recognize their
mothers’ faces, infants as young as 3 months show evidence of
integrating facial features into a whole rather than perceiving
them as individual features, and early visual experience appears
to naturally set up a neural substrate for holistic processing of
faces (Le Grand, Mondloch, Maurer, & Brent, 2004).
Why would human development favor a holistic representa-
tion of human faces rather than a feature-based representation of
human faces? One possibility stems from our earlier observation
that holistic representations facilitate recognition whereas fea-
ture-based representations facilitate recall (Wells & Hryciw,
1984). We speculate that evolutionary pressures might be re-
sponsible for mental systems favoring face recognition (and,
Volume 16—Number 1 9
Gary L. Wells and Lisa E. Hasel
CDIR 465
(BWUS CDIR 465.PDF 31-Jan-07 15:59 907858 Bytes 5 PAGES n operator=M.V.Anantha)
UNCORRECTED PROOF
hence, holistic processing) over feature-based face recall. Spe-
cifically, survival likely favored those who could readily recog-
nize faces so as to make rapid judgments of a familiar versus
unfamiliar face, friend versus foe, family versus nonfamily, and
so on. But what was the survival value of face recall? It could be
argued that face recall would have had some survival value (e.g.,
to communicate to family members the identity of a specific
dangerous individual). However, a more efficient and reliable
survival mechanism would be to simply avoid strangers, which
favors face recognition and holistic processing over face recall
and feature-based processing. Evolutionary processes that
shaped the human brain could not anticipate modern technology
and the needs and demands of a modern society.
Future research should give more weight to the question of
how people naturally process faces so as to create face-recall
systems that are congruent with natural face processing. Natural
face processing has a strong holistic bias and, therefore, systems
that require retrieval of memory for isolated facial features are
not likely to ever work well. Creating systems that reflect how
people actually process faces will require a better understanding
of what is meant by holistic face processing and why holistic
processing reduces the accessibility of information about indi-
vidual features.
Attempts to evaluate and improve face composites fit into a
larger problem in the criminal justice system. Analyses of the
first 180 DNA exonerations in the United States reveal that
mistaken eyewitness identification testimony was involved in
75% of the cases. Increasingly, it is becoming clear that errors in
human memory are accounting for more of the convictions of
innocent people than are all other causes combined. As the
historical and natural home of the science of memory, psycho-
logical science has great promise for helping to solve an age-old
problem.
Recommended Reading
Davies G.M., & Valentine T. (2006). (See References)
Frowd, C.D., Carson, D., Ness, H., Richardson, J., Morrison, L.,
McLanaghan, S. (2005). (See References)
Wells, G.L., Memon, A., & Penrod, S.D. (2006). Eyewitness evidence:
Improving its probative value. Psychological Science in the Public
Interest,7, 45–75.
Wells, G.L., & Olson, E. (2003). Eyewitness identification. Annual
Review of Psychology,54, 277–295.
Acknowledgments—This research was supported by a grant
from the National Science Foundation to the first author.
REFERENCES
Bruce, V., Ness, H., Hancock, P.J.B., Newman, C., & Rarity, J. (2002).
Four heads are better than one: Combining face composites yields
improvements in face likeness. Journal of Applied Psychology,87,
894–902.
Cooper, E.E., & Wojan, T.J. (2000). Differences in the coding of spatial
relations in face identification and basic-level object recognition.
Journal of Experimental Psychology: Learning, Memory, and
Cognition,26, 470–488.
Davies, G.M., Ellis, H.D., & Shepherd, J.W. (1978). Face recognition
accuracy as a function of mode of representation. Journal of
Applied Psychology,63, 180–187.
Davies, G.M., & Valentine, T. (2006). Facial composites: Forensic
utility and psychological research. In R.C.L. Lindsay, D.F. Ross,
J.D. Read, & M.P. Toglia (Eds.), Handbook of eyewitness psychol-
ogy Vol. 2 (pp. 59–86). Mahwah, NJ: Erlbaum.
Frowd, C.D., Carson, D., Ness, H., Richardson, J., Morrison, L., &
McLanaghan, S. (2005). A forensically valid comparison of facial
composite systems. Psychology, Crime & Law,11, 33–52.
Gibson, S., Pallares-Bejarano, A., & Solomon, C. (2003). Synthesis of
photographic quality facial composites using evolutionary algo-
rithms. In R. Harvey & J.A. Bangham (Eds.), Proceedings of the
British Machine Vision Conference 2003 (pp. 221–230). London:
British Machine Vision Association.
Hancock, P.J.B. (2000). Evolving faces from principal components.
Behavior Research Methods, Instruments and Computers,32,
327–333.
Hasel, L.E., & Wells, G.L. (2007). Catching the bad guy: Morphing
composite faces helps [electronic version]. Law and Human Be-
havior. Retrieved January 12, 2007 from www.springerlink.com.
Junkin, T. (2004). Bloodsworth: The true story of the first death row in-
mate exonerated by DNA. Chapel Hill, NC: Algonquin Books of
Chapel Hill.
Kovera, M.B., Penrod, S.D., Pappas, C., & Thill, D.L. (1997). Identifi-
cation of computer-generated facial composites. Journal of Applied
Psychology,82, 235–246.
Le Grand, R., Mondloch, C.J., Maurer, D., & Brent, H.P. (2004). Im-
pairment in holistic face processing following early visual depri-
vation. Psychological Science,15, 762–768.
McQuiston-Surrett, D., Topp, L.D., & Malpass, R.S. (in presss). Use of
facial composite systems in U.S. law enforcement agencies. Psy-
chology, Crime, and Law.
Tanaka, J.W., & Farah, M.J. (2003). The holistic representation of faces.
In M.A. Peterson & G. Rhodes. Perception of faces, objects and
scenes (pp. 53–740). Oxford: Oxford University Press.
Valentine, T. (1995). Cognitive and computational aspects of face rec-
ognition: Explorations in face space. London: Routledge.
Wells, G.L., & Hryciw, B. (1984). Memory for faces: Encoding and
retrieval operations. Memory and Cognition,12, 338–344.
10 Volume 16—Number 1
Facial Composite Production by Eyewitnesses
CDIR 465
(BWUS CDIR 465.PDF 31-Jan-07 15:59 907858 Bytes 5 PAGES n operator=M.V.Anantha)
Author Query Form
_______________________________________________________
_______________________________________________________
Dear Author,
During the copy
-
editing of your paper, the following queries arose. Please respond to these by marking up your proofs with the necessary
changes/additions. Please write your answers clearly on the query sheet if there is insufficient space on the page proofs. If returning the
proof by fax do not write too close to the paper's edge. Please remember that illegible mark
-
ups may delay publication.
Journal
CDIR
Article
465
Query No.
Description
Author Response
.
No Queries
