Washington
24 January 2008
Researchers say they have improved the accuracy of automatic face recognition systems, which are being used more and more by crime prevention and national security agencies. The investigators have created what they call an "average face" that, they say, allows the machines to make a correct identification in every case. VOA's Jessica Berman explains.
Experts say that automatic face recognition systems, in which a computer attempts to recognize an individual by checking his or her face against thousands of images stored in a database, are gaining in popularity, but are not very reliable.
There won't be a match if the person's face is not contained in the database, or a match may be overlooked if a picture that is in the database looks sufficiently different than the individual standing in front of the system's electronic eye.
Experts say that's because the stored photograph may be several years old and bear little resemblence to the individual, or the pose or light may be different.
Psychologist Rob Jenkins of the University of Glascow and his colleague Mike Burton believe they solved the problem.
They randomly gathered 20 images of 25 famous male celebrities on the internet and fed the photos into a face recognition system that had a database of 31-thousand photos of famous faces.
The recognition system accurately identified the individuals about half of the time .
The researchers then created what they call an "average face" of each celebrity by taking his images and averaging them out. They did that by removing "unhelpful" aspects of the images, such as lighting effects, and highlighting structural features of the face, Jenkins said.
"So, what we were able to show is that if you take these highly variable images, and apply the averaging process to them, then the recognition performance goes right up. In the study we reported, it goes from 54 percent, which is the performance that it got on standard photographs, up to 100 percent. So it's a really huge leap," she said.
Jenkins sees no reason why "average face" photos could not be made of everyone and used for identification purposes. He says the average faces could also be used with existing face recognition technology for use, say, at airport check points.
To be sure, the Scottish psychologist is ready. He made an "average face" image of himself using 20 different images. "What's striking to me is how bad a likeness my passport photo is and how a good a likeness the average image is. Certainly, think if I were at customs deciding whether to admit myself or not, I think I would have to say 'no,'" he said.
Jenkins and Burton report their work in the journal Science.