IT IS an inevitable part of any crime report: a grainy composite 鈥減hotofit鈥� image of a dubious character that purports to resemble the criminal. But how do the police come up with these pictures, and how accurate are they?
The field has come a long way. Before the 1970s, a police artist would draw a suspect鈥檚 face straight from the witness鈥檚 verbal description. The early composite systems used libraries of separate line-drawn facial features such as hair, eyes, nose and mouth that the witness could choose from to build a suspect鈥檚 face. Then in the early 1970s Jacques Penry, a French photographer, developed a new composite system that used facial features taken from photographs of real faces.
The results of these composite systems were very disappointing, however. The images rarely resembled the face they were meant to represent. Moreover, they were not much more realistic even when the operator had the actual face in front of them. The biggest problems were the restricted range of facial features available, the limited opportunity for artistically embellishing the image, and the fact that it was difficult to move the facial features around to change the way they fitted together, known to be extremely important for facial recall.
Advertisement
Most of these shortcomings have been overcome in the computer-based composite systems used today, such as E-Fit and PRO-fit. These work in much the same way as the manual kits, with skilled operators working with witnesses to recollect features, but they are much better at placing, blending and embellishing the features. It is also now possible to show angled composites as well as full-face ones. As Hayley Ness at the University of Stirling, UK, has shown, using a combination of these composites can make it easier to remember a face.
Nonetheless, witnesses still find it difficult to produce good likenesses of suspects from memory even with the new electronic composite systems. At the University of Stirling, UK, we have been trying a different approach. We reasoned that we might end up with a more accurate image by pooling together the memories of different witnesses and creating an average. The idea was that this would improve the likenesses and reduce the errors, since genuine features would probably be remembered by a number of witnesses and would therefore be enhanced by an average, while the errors would be more random and would be averaged out. We used a computer program to morph together composites produced by different witnesses, and found that they were generally better than or at least as good as the best of the individually produced composites.
Even so, composites generate identifiable resemblance only in about 20 per cent of cases. Why are good likenesses so hard to produce? The problem is not with the technology, since unlike the old systems, operators can produce a good likeness when they have the actual face in front of them. Instead, the difficulty lies with the way our brains remember and recall faces. We tend to perceive and remember faces as wholes and in terms of how the facial features relate to each other, rather than the separate parts in isolation. Asking witnesses to describe parts of the face separately seems at odds with this.
Charlie Frowd and Peter Hancock, also at Stirling, have been developing a new system of composite construction called EvoFIT, based on this holistic approach to face perception. The witness gives a general description of the suspect and is then shown 18 synthetic faces that resemble the description (the faces are synthetic so that no innocent person risks being framed). From this set of 18 the witness chooses and ranks the six faces closest to their memory of the suspect. EvoFIT then uses a genetic algorithm to 鈥渆volve鈥� a new set of 18 faces. The witness then selects the best set from these, and the process continues until the system evolves a likeness that the witness thinks is satisfactory.
The researchers have recently extended the system to show faces from various angles. The only feature that EvoFIT requires witnesses to remember in isolation is hairstyle, which for technical reasons cannot easily be evolved. EvoFIT鈥檚 advantage is that it is based almost entirely on recognition of whole faces, and in recent trials it has often outperformed other systems such as PRO-fit. This is encouraging, as the range of faces and hairstyles within EvoFIT is still rather limited.
But is all this painstaking work really necessary now that many crimes are committed in full view of CCTV cameras? Research carried out in the UK at the universities of Glasgow and Stirling has demonstrated that people can recognise familiar faces even in low-quality CCTV images. However, the research has also shown that the crucial trigger is the person鈥檚 face, and if the face is concealed from view people find it very hard to identify even very familiar people. In addition, some crimes will always be committed away from CCTV cameras, such as in the home. In many investigations, making the most of human memories will often be our best bet.