Forensic composite imaging

     Law enforcement has used facial compositing to help solve crimes since Billy the Kid was a fugitive in the American Wild West.

     Conventionally, sketch artists created hand-drawn renderings with charcoal or pencil. But today the use of sketch artists is becoming increasingly rare in criminal investigations, often reserved for only the largest departments. The sketch artists of old are quickly being replaced by facial composite software.

     While the work of artists is being taken over by computerized tools, the functions of forensic composite imagery have expanded. It is used for identification of persons or objects, image modification, depictions of crime scenes, and reconstruction and post mortem identifications, says James Lucas, adjunct faculty at Oakton Community College of Illinois.

     But some experts are questioning the accuracy of these tools — and of composites or photofits in general.

     In an article appearing in the February issue of "Current Directions in Psychological Science," Gary Wells, professor of psychology at Iowa State University whose studies of eyewitness memory are widely known and cited, and co-author Lisa Hasels, also from Iowa State University, point to several studies that indicate facial composite systems produce a poor likeness of the intended face. According to the article, researchers had inferior results in studies where individuals attempted to create composites of famous celebrities. In one study, just 2.8 percent of participants correctly named the likeness of a well-known celebrity created with face composite software.

     Researchers like Wells are concerned because of the potentially negative consequences of a bad composite. Broadly distributed to law enforcement, newspapers and the public, an inexact image may lead an investigation astray and produce wrongful convictions. A sketch of the Unabomber in a hood and sunglasses created a virtually useless rendition of Ted Kaczynski, and he eluded law enforcement for many years. On the other hand, strong links tied Timothy McVeigh to the Oklahoma City bombing, among them the fact that his appearance matched eyewitness accounts used by FBI artist Raymond Rozycki to develop a composite drawing.

     "There is a presumption — if we have a tool use it. But [what we're saying is] wait a minute, be careful here," Wells cautions.

Flaws in thefacial composite process
     Dr. Harry Wechsler, a noted expert on biometrics and face recognition, who serves as professor of computer science and director of the Center for Distributed and Intelligent Computation at George Mason University in Fairfax, Virginia, points to articles by G. Daniel Lassiter published in "Current Directions in Psychological Science" (2002) and Dr. Charlie Frowd, a researcher at University of Central Lancashire, whose primary research involves improving the quality of facial composites, in "Psychology Crime and Law" (2005).

     Frowd's article titled "Contemporary composite techniques: The impact of a forensically relevant target delay" outlines the lackluster performance of composite naming using a two-day delay between the time a witness saw the perpetrator and when he or she worked with the forensic artist. His study concludes that composite naming, that is, identifying a subject by matching a composite against a database, though a photo lineup or a police bulletin, "was surprisingly low," says Wechsler. In fact, composite naming was successful in only 3- to 8-percent of the studied cases.

     Yet facial composite software may not be to blame, Wechsler says. Numerous factors affect the accuracy of eyewitness composites: A delay following the event, exposure time to the subject, target distinctiveness, emotion, stress, all play an important role in the composite produced.

     Whether manual or computerized, composites depend on verbal descriptions that are not always accurate. A basic problem, Wechsler says, is that photofits mostly consist of shape and very little texture. These contours are embellished with scant skin texture, and lack other details that make a person recognizable.

     In addition, researchers find issue with the method in which eyewitness composites are assembled — by individual features.

     Psychological research has shown that people store faces in memory in a holistic way, says Wells, whose research on eyewitness identification is funded by the National Science Foundation. "They don't store individual features but an overall holistic image of the face. And if faces are stored holistically, the best retrieval is holistic," he says. "[Witnesses] don't really have good retrieval access to individual features. Any task to retrieve individual features is going to be profoundly difficult."

     Knowledge of holistic processing may be a boon to future face composite systems that will utilize "whole-face" methods for face recall, according to a February 11, 2007, article in "Science Daily," for which Wells was consulted. Such systems begin by generating a random set of faces. The witness then selects the face most similar to their memory of the perpetrator. This will be the "parent" face that yields a set of similar looking faces, which are the result of several mutations to the parent face. The witness again makes a choice and the process continues until it's impossible to discriminate between the options from his or her memory of the perpetrator.

     EvoFIT software, with British firm ABM as its industrial partner, is one facial composite system based on a holistic face coding scheme and an evolutionary interface. Using this system, witnesses choose from a selection of faces that bear a resemblance to an assailant (a composite is evolved over time by breeding together the selected faces). In recent experiments by Frowd, EvoFIT has outperformed other current composite systems (in the most recent realistic study, EvoFIT reached a level of naming roughly twice that of another United Kingdom composite system).

     Frowd, who has conducted extensive research in the area of facial composites, along with principal investigator Dr. P. Hancock, is currently working on a U.K. government grant at the University of Stirling, Scotland. The aim of the project, called Evolving a Better Composite, reports, is to develop and exploit the holistic (whole face) nature of EvoFIT. There is good evidence that composites produced by EvoFIT are better than other U.K. systems, but research is now required to optimize performance, notes Frowd. Currently, the system generates faces with random characteristics, for a given sex and race, and the user selects those that most resemble the target.

     Researchers wish to make the system more user-friendly by identifying psychologically useful variables, such as age and obesity, and implementing these within the model. This would allow a witness to request that a face be made older, for example, something which requires considerable artistic skill with current facial composite systems, explains Frowd.

     A second aspect of the project is to explore better ways to help witnesses remember relevant details. Current police interviewing techniques concentrate on facial features (e.g. describing the nose), which may work well for today's feature-based composite systems, but gets in the way of remembering what the face as a whole looked like. The researchers' aim is to develop interviewing methods that better match the holistic nature of EvoFIT, and to predict that a combination of these approaches should also enable witnesses with a limited recall (of an assailant) to construct useful composites.

Computerized vs. manual composites
     Frowd et al. also conducted a study titled "An evaluation of U.S. systems for facial composite production" that evaluated the reliability of two systems:

  • San Mateo, California-based Wherify Wireless Inc.'s Faces 3.0, and
  • Scottsdale, Arizona-based Identi-Kit Solutions' Identi-Kit 2000.

     The study, which is to be published later in 2007 in "Ergonomics" is posted on

     In this study, Frowd states researchers found U.K. software programs produce composites that are correctly named about 20 percent of the time when participant-witnesses attempt construction either immediately or a few hours after seeing a target face. Unfortunately, he says, when participant-witnesses are required to wait two days before they can construct the composite — a situation typical of real police witnesses — composite naming fell to a few percent correct at best, according to statistics gathered in research for journal articles "The Process of Facial Composite Production" in "Forensic Psychology and Law" and "Parallel Approaches to Composite Production" in "Ergonomics."

     A comparison of three U.K. programs, a sketch artist and Faces 3.0 (a program not differing so much from later version 4.0 that results are invalid) were the subject of another Frowd study. It found that composites done by a sketch artist were correctly named 8 percent of the time, but the other systems were worse — naming was less than 4 percent.

     A 2005 study by Frowd et al. comparing computerized systems with manual systems found that all other composite systems tested were inferior to a sketch artist. While it may be that the skill of the person was responsible for this, he says the mode of representation might be as well. It turns out that sketch artists working with witnesses tend to produce composites with less shading compared with systems, such as ABM's PRO-fit and FACES, which utilize photographed features, especially in areas around the forehead, cheeks, nose and chin. It might be that these features, Frowd notes, are better left blank rather than including potentially misleading shading information. As such, sketched composites may contain less incorrect information and may be more identifiable.

More research needed
     More research is definitely called for. Anecdotal evidence touting the successes of composite sketches can be dangerous. "It's in the nature of anecdotes to be selective and to ignore all cases where it didn't work," says Wells. "Our approach has been to drop the anecdotes — we like to study in controlled conditions … we know which face [witnesses] saw and how well they do in generating the face. Software has improved incredibly but there's still a problem in getting likenesses. The [issue] doesn't seem to be within the software itself."

     Wells calls Frowd's research "promising," and continues to hope for more controlled studies and police software programs that stand up to scientific criteria. "If we can't make them better," he says, "we should use them less or as a last resort."

     On the other hand, taken with other evidence, composites can assist investigators. "Photofits factor out many biometric features," says Wechsler, "narrowing down the police work to identify suspects and thus making [their job] easier and more efficient." Efficiency increases because investigators narrow the subset of suspects, and accuracy improves because they can spend more time on a reduced set of suspects.

     Perhaps Lee Farrow, FACES global account manager at Wherify Wireless Inc., says it best: If we compare finding a suspect through use of facial composites to finding a needle in a haystack, "facial composites may make the haystack a little smaller or the needle a little shinier."

     She continues it's critical to remember that a facial composite gives investigators "a direction but it's not the nail in a coffin — there's additional evidence like DNA or fingerprints." It narrows down the suspect, but "it's not the finalized tool to arrest someone."

Donna Rogers is a freelance writer based in Huntington, New York. She can be reached at