London, Jan 25 : Averages of several snapshots could not only improve ID card security, but also allow facial recognition software to spot familiar faces as well humans.
Inspired by the way people recognise faces, psychologists Rob Jenkins and Mike Burton of the University of Glasgow, UK, found that we are much better at recognizing averaged faces than if shown an individual photo.
They said that as we see a person from different angles and in varied lighting, we make a mental image of their face that averages out all those experiences. They further showed that averaging multiple photos of a person can also help facial recognition software work better.
To prove their findings, the researchers plugged individual photos of male celebrities into an online genealogy service, MyHeritage.com.
This website allows people to compare their photos against a large database of celebrity images to see which famous people they most closely resemble. The site is powered by a facial recognition program called FaceVACS.
The duo supplied in photos of 25 different male celebrities, right from Bill Clinton to Jack Nicholson, in order to find out if MyHeritage would correctly identify them.
MyHeritage pinpointed the celebrity only 54pct of the time, out of more than 450 photos. However, when the researchers made a composite photo for each celebrity using several photos, MyHeritage identified them every time.
The researchers decided to make things a bit harder for the recognition software by excluding the 54pct of photos that had a correct match in the first test, in case the celebrities were easier to recognize in those particular photos. Then they combined the remaining photos into composites and fed those into MyHeritage. This resulted in positive matches 80pct of the time.
Jenkins said that the site's results with averaged photos match the performance of humans asked to recognise familiar faces.
"This is the first time software has shown anything like this level of accuracy for photos that are not controlled [for pose or lighting]. The good thing about image averaging is it washes out all the things like lighting and pose that differ between photos. You're extracting the essence of a person's appearance. Although some information is lost, "you're only throwing away bad information," he said.
The averaging technique first identifies a face's landmarks, such as the corners of the eyes or the mouth, to align several photos. If photos show a person at different angles, the software stretches the image so that landmarks on different photos align with each other. Finally the photos are merged together.
Researchers said that including an averaged image on passports or other identity documents could make computerised security screening, for example in airports, better. An averaged composite of several photos taken over the last 15 years, however, is a much closer match.
"My passport photo looks nothing like me," Jenkins said.