New photofit 'evolves' a suspect's face
- 10:00 19 March 2005
- Exclusive from New Scientist Print Edition. Subscribe and get 4 free issues.
- Paul Marks
Half an hour after being mugged, the victim is again staring her attacker in the face. But the assailant has not returned. What the victim is looking at is an image on a police laptop running software that can "evolve" a realistic likeness within minutes, while her memory is still fresh.
This novel photofit system was unveiled in London, UK, last week at a conference on crime-fighting technology. The computerised photofit systems police now use store hundreds of variations of facial features, such as eyes, noses and chins, in a database.
A police officer interviews the witness and attempts to piece together a likeness. "They will look through a gallery of about 150 eyes and put the right ones, pastiche fashion, into a face. But witnesses find this a bit overwhelming," says Chris Solomon of the University of Kent in the UK. The procedure can take hours and the resulting composite is not very lifelike.
So Solomon's team, and Peter Hancock and Charlie Frowd at the University of Stirling in Scotland, are working independently on better ways to produce photofits using genetic algorithms. Engineers often use this type of software to come up with optimal designs for, say, Formula 1 cars, or the perfect mixture for an alloy. A genetic algorithm takes possible solutions to a problem and repeatedly "evolves" them until it finds the one that works best.
Kent's EigenFit software and Stirling's EvoFit package work in similar ways. Based on the sex, race and hairstyle of the person the witness remembers, the computer produces nine random faces, from which the witness chooses the one that seems the closest likeness. The algorithm then uses this face to mutate a new set of variants.
The cycle continues until the witness is happy with the likeness. Each generation can be calculated in seconds, making the process far quicker than retrieving facial features from databases and trying them one by one.
But just what is being evolved? Each face is represented by an array of 50 numbers called principal components. "If we change just one of these parameters it alters the face, albeit rather subtly," Solomon says. "It might make the skin a bit darker, more wrinkly, or move the nose up the face a bit." Once a feature, say the mouth, is correct it can be "locked", and the rest of the face evolved around it.
In early tests, volunteers were about twice as likely to recognise a face constructed through the algorithm-based software as through today's photofit mugshots, Solomon says. And when police in Northampton asked a victim of a crime to construct an image with EvoFit, "she was astounded at the quality of the likeness to her attacker", senior investigating officer Paul Spick told New Scientist. "The holistic way it produces a whole face is very impressive."