Scientists have developed a ”face averaging” technique which dramatically improves the success rate of computer face-recognition systems and may be used to streamline airport security, solve crimes captured on CCTV — and find celebrity lookalikes.
Using the new technique boosted the performance of a widely used face recognition software package from 54% accuracy to 100%, University of Glasgow researchers said on Thursday in a paper published in Science.
The creator, Rob Jenkins, said the idea was based on the fact that people are able to recognise familiar faces much better than new ones. ”If I just show you two photographs and say, ‘is this the same person as that?’, we are actually pretty bad at that. We are actually much worse than we think we are unless we are familiar with the face,” he said.
However, we are much better at recognising the faces of people we know. By making an average of 10 images of the same person his software is able to eliminate variation from, for example, different lighting or camera angles.
”These things affect the image a great deal but they don’t tell you anything about who it is,” he said. ”It is like you are extracting the essence of that person’s face.”
Automatic systems would be very useful for recognising wanted criminals automatically on CCTV or for making check-in at airports more streamlined, but at present they typically make too many mistakes except under very controlled conditions.
Jenkins and his colleague Professor Mike Burton tested the averaging approach using FaceVACS, a system that is being tested at Sydney airport.
The website MyHeritage.com uses the software in a celebrity lookalike service. Surfers submit images of themselves to the site, which matches them to the nearest celebrity picture on its database of more than 31 000 photographs. Jenkins and Burton submitted images of 459 celebrities they knew were on the database. The system matched them to the correct celebrity 54% of the time. When the pair created average celebrity faces from 10 images and resubmitted them to the website, the software was correct 100% of the time. ”That’s the first time anyone’s reported anything like that level of accuracy on such a variable set of images,” said Jenkins.
To make the test even tougher, they created averages using only those images that the software had been unable to match in the first test. In this test the software recognised the averaged faces correctly 80% of the time. ”We have fixed the baseline at zero, so any improvement we can attribute just to the averaging process,” he said.
Jenkins said his approach did not compete with current technology. ”One thing I really like about this approach is we are proposing a technique, not a device,” he said. ”We are saying, you can keep the same machines, but if you can change the input you can radically improve their performance.” – Guardian Newspapers Limited 2008