"Whole face" software to improve composite sketches of criminals

The criminal sketch artist is one of the fondest clichés of the police television drama. An old lady is mugged on the street and sits for hours with a patient pencil and watercolorist, dutifully instructing him to make the chin sharper or the eyes darker. Nowadays, of course, this is more often done with computers rather than by hand, but the idea is the same: the rapid dissemination of a image purporting to be the perpetrator is one of the best ways to solve a crime. It is just too bad, researchers at Iowa State University say, that it rarely works.

According to professors Gary Wells and Lisa Hasels, studies have shown that so-called facial composite systems produce almost useless likenesses. In one particular study, celebrities were described by participants and then drawn using software programs. Among the other participants chosen to identify the drawing afterwards, only 2.8 percent succeeded in doing so. In a different study, participants failed to identify composites of their own classmates. Having reviewed the vast literature on the topic, Wells and Hasels say the problem is not a particular weakness in the software but a failure to understand the difference between how humans remember faces and how the composites are actually produced.

“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,” said Wells. This means that humans use complex representations such as multidimensional similarity to other faces or relative sizes and distances of features when recognizing another person. The process of facial composition, however, does not take this into account. One solution, however, is already in the works. Wells and Hasels point to promising development in software that relies on “whole-face” methods for face recall.

The approach reminds us of an emerging IT security technology that asks the user to select a pattern of predetermined faces from a grid before accessing his computer. The premise is that faces are easier to recall than numbers. In the whole face method, the system presents the witness with a random set of faces from which the witness chooses the most familiar. Then another set is generated, and the witness begins whittling down the options until a face closely approximating the suspect is identified.

-read more in this university news release