Computer morphing e-fit faces created by several witnesses to a crime together can increase the accuracy of the final image, new experiments show.
Peter Hancock at the University of Stirling, Scotland used software to blend four versions of a famous face, each created by a 鈥渨itness鈥�. When subjects were given a list of six names to choose from, they recognised the composite 93 per cent of the time. Recognition of single images varied from 53 to 78 percent.
鈥淭he idea behind it is that each face captures some aspect of the real face,鈥� says Hancock. 鈥淏ut adding them all together should allow the common aspects to come through and the independent variations to cancel out.鈥�
Advertisement
Professor David Perrett from the University of Saint Andrews told 快猫短视频: 鈥淚t鈥檚 very clever.鈥� He says the logic of averaging out 鈥渘oise鈥� from witnesses who may have only briefly seen a face is sound. 鈥淚t goes back 100 years to Galton鈥檚 work in which he tried to create an average criminal face.鈥�
However, a change in UK police regulations is needed if morphing is to be used. Currently information from different witnesses cannot be combined because if the witness is required in court they can only present their own evidence.
鈥淥ffspring鈥� faces
Hancock is also working on a system of evolving faces, to allow an individual to create a better e-fit. He says Pro-fit, currently used by police to create images does not work that well because we do not recognise faces by assembling individual features.
His system, Evofit, creates random variations of a witness鈥檚 first attempt to create 鈥渙ffspring鈥� faces. The witness chooses the best of these and repeats the process, perhaps four times, until the best images has evolved.
Evofit has been used once in an investigation to capture a serious sexual offender in Northhamptonshire. The final image was shown on the UK television programme, Crimewatch, and generated a large number of leads on individuals.
However, laboratory experiments with Evofit show lower recognition rates than with Pro-fit. Hancock thinks improvements are possible in the underlying face model and also in how the faces are presented. One concern is 鈥渕emory interference鈥� 鈥� that the showing of too many evolved faces can confuse the witness.
Future work may involve moving to 3D images and extending the database of faces beyond young adult white male faces 鈥� 鈥渢hough most crime in UK is committed by this group鈥� notes Hancock. The faces currently in the database are all of policemen, a nice counterpoint to the old photofit system which was entirely based on criminal faces.
Hancock presented his research at the British Association鈥檚 Festival of Science in Glasgow.