The problem with the traditional “photofit” process is that the human brain recognizes faces holistically, not as a collection of isolated features.

In the Dark Ages of policing, an eye-witness to a crime would be asked to reconstruct the perpetrator’s likeness, jigsaw-like, from a daunting selection of isolated eyebrows, noses, and haircuts. This low-tech process tended to produce bizarre and inaccurate results.

This mismatch between human psychology and crimefighting methods may have been resolved by two research groups in the U.K. who have independently stumbled upon similar solutions as to how to produce a better “facial composite” of a suspect in a more psychologically natural manner. Peter Hancock and Charlie Frowd at the University of Stirling, Scotland, have produced a system called EvoFIT,1,2 while Chris Solomon at the University of Kent, England, has developed a similar system called EigenFIT.3 Now the race is on to commercialize both these systems and win the approval of the police and security services.

So what is the problem with the traditional photofit process? Charlie Frowd from the Scottish team explains: “Peter knew that the normal method used to make composites does not work very well,” he says, “we are not good at describing and selecting individual facial features, but are very good at selecting whole faces which look like someone we’ve seen.”

Utilizing the fact that humans are better at whole-face recognition than cherry-picking of features, both systems use a computational search technique known as a genetic algorithm to “breed” together several faces that are judged to be similar to that of the suspect. Over several “generations,” the composite should approach optimal similarity to the suspect.

A genetic algorithm adopts concepts borrowed from evolutionary biology to “evolve” an optimal solution to a problem. The technique works by starting with an initial population, the search space, which can be randomly generated or narrowly specified. The Darwinian “fitness” of the individuals in the population is evaluated according to some criteria and the highest-ranked individuals are selected for reproduction. A new generation is produced from these “parents” through combination of their features and/or addition of random features (mutation). The fitness of the resulting “offspring” is then evaluated. The whole initial population, or just the worst-ranked individuals, are replaced with the offspring, and from that new population new parents are chosen to breed. This process continues until some terminating condition is reached.

For example, an EvoFIT session goes like this: From a randomly-generated selection of eighteen faces, a witness is asked to choose the six faces that most resemble the suspect. These six faces become the parents of eighteen offspring faces generated by combining the features of the parents. The witness then chooses another six from the offspring population to become parents, and so on. The “features” being selected and mutated are values of around fifty “principle components” that describe the structure of the face. The process can proceed indefinitely, but usually produces a likeness acceptable to the witness in about four generations, as long as the initial selection (or “search space”) was adequate. This likeness can be saved and embellished with haircuts and clothing using normal photo editing software. The system interfaces with existing computerized composite systems such as E-Fit, so that their pre-existing databases of hairstyles, piercings, scars, and other facial adornments can be utilized.

One drawback of genetic algorithms generally, is that they can converge on the wrong solution if the initial population is poorly chosen. “This may happen,” Frowd admits, “but if it does, the system is rolled back and started again.” Project leader Peter Hancock elaborates, “We suspect the quality of the end result depends rather on there being something reasonable at the start,” he explains, “hence our attempts to narrow down the starting population.” When asked how well the initial population needs to be defined, he replies, “ not very much.” The techniqueis remarkably robust. Tinkering with the features of the randomly-generated initial population doesn’t greatly improve performance, “beyond general age, race, and sex,” he concludes.

EvoFIT has also been tested experimentally. Participants are asked to reconstruct a random face that they have seen hours or days before hand. The system performs well in comparison to existing conventional systems, but with the additional appeal of its intuitive interface and ease of use.

The Stirling team is now aiming to improve EvoFIT’s ergonomics, or user-friendliness. “We have designed a set of tools which allow a witness to manipulate the overall appearance of their composite face. This includes age, masculinity, facial weight, etc. It works very well!” says Frowd. Previously, if a witness requested that a composite be made to look older, the image would have to be altered by a skilled forensic artist. However, the rules regarding how the face and the shape of the head change with age can easily be simulated. “Age is easy. We have an average young man and an average older one and can simply apply the transform to age a given face, or the population as a whole,” says Dr. Hancock.

EigenFit composite imges

Producing a fit takes a bit longer than with previous traditional systems (“Maybe 15-30 minutes longer,” says Frowd), which could be considered a problem in a criminal investigation where time is of the utmost essence. But the benefits of the system quite clearly outweigh this issue. “I’ve never really been worried about that,” argues Frowd, “as the composite is more identifiable than from the other systems and so the extra time is worth it.”

The EvoFIT project sprang naturally from the face recognition research for which Stirling has become famous. Hancock wasn’t initially looking to fight crime: “I realized that I had the tools needed from my previous work with faces and genetic algorithms.” Such serendipitous discoveries can only emphasize the importance of pure research.

The differences between EvoFIT and the English system, EigenFIT, are subtle, but they center on the exact nature of the genetic algorithm being implemented and how the facial model is constructed and manipulated. “The principles are broadly similar,” says Chris Solomon, but compared to existing photofit systems, “we believe that EigenFIT has been better adapted to human use and cognitive processes.” EigenFIT requires the witness to select faces based on their similarity to the target face. The face that is the current front runner (or stallion, as the terminology amusingly puts it) is then either bred with an entirely new face or is “cloned” and each clone randomly mutated. The witness then selects the new faces and if one is judged to be a better match than the stallion, the old stallion is retired in favor of the new.

Solomon argues that EigenFIT is faster than conventional photofit systems: “Time is money for the police just as for any other organization. Our studies have shown that the procedure takes an average of under half the time that a conventional composite takes.” He continues, “We view this as a very important featureespecially in the context of volume crime.” This speed difference could be the advantage that the southern stallion needs to pull ahead of the pack.

Another advantage of the EigenFIT system is its ability to manipulate faces along subjective dimensions. EigenFIT uses a “face space” that is a statistical model of many real faces. These initial faces can be rated on any subjective characteristic such as kindness, masculinity, or attractiveness. Using these subjective ratings of the component faces, any face can be manipulated automatically to look more feminine or more aggressive. Solomon is clearly pleased with this aspect of the system. “Currently, at least, EigenFIT has a…comprehensive facial model, which allows a greater control over the construction process,” he says.

However, isn’t there the potential in both systems to bias witnesses’ memoriesby showing them lots of faces? “We can’t easily define that, but inevitably there will be some contamination from seeing lots of faces,” responds Hancock. “But that’s also there for witnesses who use the other systems,” adds Frowd. “ We’ve some data which shows that there is no more interference (if any) between an EvoFIT user and a user of another system,” he concludes.

Thanks to their conceptual simplicity and their ease of implementation, genetic algorithms have found uses in various disciplines including physics, economics, chemistry, and computing. These evolutionary algorithms may one day help to produce evolvable hardware, a Swiss-army-knife style black box that can be plugged in to other devices, without needing to be explicitly programmed to carry out any particular job. The device would only need feedback as to how well (or badly) it is performing the task it has been assigned. This plug-and-play nirvana may be a long way off, but EvoFIT and EigenFIT are already gaining acceptance in the real world.

EvoFIT has been used by Northamptonshire Police in the hunt for the “Beast of Bozeat,” a rapist who evaded capture for seven years. Although DNA evidence, rather than the EvoFIT facial composite, eventually led to the conviction of nightclub bouncer James Davies, EvoFIT is now being considered by other police forces. “We’ve trained an officer in Mid Wales. Next month we’ll start training Lancashire police,” says Frowd. “There’s interest from several other forces in the UK.”

With an injection of funding from the Engineering and Physical Sciences Research Council, EigenFIT has been rapidly commercialized.4 “It has been used in a substantial number of cases now,” says Chris Solomon. He continues: “it has now been commercialized under the name EFIT-V and is being used by several U.K. police forces.”

The psychological theory behind both EigenFIT and EvoFIT is very similar, as is the software implementation. It remains to be seen whether the police and forensics community will embrace this revolution inspired by evolution.


  1. Frowd, C.D. (2002) EvoFIT: A Holistic, Evolutionary Facial Imaging System, Unpublished PhD thesis, University of Stirling, 2002.
  2. Frowd, C.D., Hancock, P.J.B., & Carson, D. (2004) EvoFIT: A holistic, evolutionary facial imaging technique for creating composites. ACM Transactions on Applied Psychology (TAP), 1, 1-21.
  3. Gibson, S., Solomon, C., & Pallares-Bejarano, A. (2003) Synthesis of photographic quality facial composites using evolutionary algorithms. In: Harvey, R., Bangham, J.A. (Eds.), British Machine Vision Conference, British Machine Vision Association, 9th-11th September 2003. BMVA Press, University of East Anglia, Norwich, UK, pp. 221-230.
  4. EPSRC Press Release (2006) “EigenFIT - Fighting Crime Effectively” Published 21 Aug 2007. EPSRC. Accessed 12 Oct 2007.


Craig Aaen Stockdale, Ph.D., is a postdoctoral research fellow for McGill Vision Research, Department of Ophthalmology, McGill University, Montreal, QC, Canada. Craig can be reached at