Jan. 24, 2008 -- Humans are fantastic at recognizing familiar faces. Computers, on the other hand, are pretty terrible at it.
Now researchers have developed an imaging technique that merges several photos of the same person into one composite. When composites of different people were shown to a computer, it was able recognize all of them.
Such a method could improve face recognition programs already in use to identify suspects at airports, for example, or as part of immigration and employment verification.
"Keep the machines that you're using but give them a different input, and that will improve their performance radically," said Rob Jenkins of the University of Glasgow.
Jenkins and professor Mike Burton published their research in today's issue of Science.
Jenkins and Burton stumbled upon the technique as part of their psychology research. They were studying the fact that although humans are great at recognizing photographed images of people they know, they stink at picking out photos of people they don't know.
In this regard, humans and computers struggle with the same issues. Varied light conditions, expressions, unusual face angles and diverse ages all conspire against positive recognition. In short, ten images of one stranger can look like ten photos of ten different people.
But when the researchers morphed several photos of the same person, the averaged look was easier for people to discern when comparing it to another image of the person. Jenkins and Burton theorized that it might be easier for a computer to make the connection as well.
So they tested their technique using the public Web site, MyHeritage.com, which runs an industry standard recognition software called FaceVACS. The site contains more than 31,000 photos of celebrities. Users can upload images of their faces and let FaceVACS pick out the celebrity they most resemble.
To gauge the general accuracy of FaceVACS, Jenkins and Burton submitted photos of celebrities known to be in the MyHeritage database. Of 459 images, the program correctly identified about half.
But when the researchers used those photos to create composites of the celebrities and then submitted those to the Web site, FaceVACS correctly identified every one.
"The performance shot up to 100 percent," said Jenkins. "This is a huge leap in a field where people get quite excited about improvements of 2 or 3 percent."
"It's a new approach that not many people have studied," said Jonathon Phillips, an expert in biometric and automatic face recognition systems at the National Institute of Standards and Technology in Gaithersburg, Md.
But, he pointed out, the morphing technique is not yet automated. Each image must be manually converted to grayscale, rotated, resized and cropped so that critical points on the faces, such as mouth and eye corners line up. The morphing and recognizing processes happen separately.
"The question is how do you combine them into an effective technique?" said Phillips.
Jenkins and Burton would like to see their method improved upon, but are going back to human experiments to tease out the range of differences people exhibit with their recognition skills.
Related Links:
our sites
video
mobile
shop
stay connected
corporate
