Researchers from the University of Glasgow, led by Dr Rob Jenkins and Professor Mike Burton, have developed a new technology that drastically improves the success of both human and automated systems at matching a face to a photo. Dr Jenkins spoke about this work at the BA Festival of Science on Tuesday at The BA Joseph Lister Award Lecture.
The team worked on the principle that the more familiar a person is with the face in question, the easier it is for them to recognise the face in a photograph. They collected different images of each person’s face and averaged them to make one image. Only 10 – 11 photos were required to stabilise the image, with any extra images after that making little difference. The average face image was also surprisingly resilient to errors, even with the ‘contamination’ of a third of the faces being of other people, the average face hardly changed.
Dr Jenkins said: ‘The resulting images are quite uncanny, seeming to bring out the true essence of each face.’
The averaged faces were then checked against the individual faces from which they were made. Both humans and machines were significantly better at recognising the average picture. Jenkins explained: ‘This is because the averaging process washes out aspects of the image that are unhelpful, such as lighting effects, while consolidating aspects of the image that are diagnostic of identity, such as the physical structure of the face.’
With the proposed identity card scheme looming, the breakthrough of better-than-photo recognition accuracy raises the question of whether face databases and ID documents should contain identity averages, rather than standard photographs.
‘This boost in face recognition accuracy has major implications for crime prevention and national security policies. It also demonstrates that with face recognition, as with so many other problems, we can improve machine performance by mimicking nature's solution,' said Dr Jenkins.