Jan. 24, 2008 — -- Researchers claim they have perfected a system that uses computers to accurately identify images of people's faces, which could aid in the apprehension of criminals in public places such as airports that use surveillance cameras, according to a study released Thursday.
But some experts still doubt that facial recognition software can be used to accurately pick people out of crowded, public areas.
Comparing a database of images of criminals, to a real live person in a crowd, has been very difficult, concedes Rob Jenkins, a professor in the psychology department of the University of Glasgow and co-author of the study released in the journal Science.
But using a newly developed program at the university, computers were found to be 100 percent accurate when using what they call an "averaged" face image, made up of 20 photos, Jenkins and co-author Mike Burton wrote in the paper.
"The great thing about this averaging process is it just washes out all these differences of single photographs. The lighting and the pose all kind of becomes neutralized," Jenkins told ABCNEWS.com. And what you're just left with is the core of the face. The aspects of the image are consistent from one photo to the next."
Facial recognition programs have been used for years. The most successful applications have been in the government or the private sector, mostly to help identify employees seeking access to sensitive areas. Casinos have also been using the software to help spot criminals or known card cheats sitting at gaming tables.
That's a whole lot different then picking random people out of a moving crowd and matching the images to a datatbase of known criminals or terrorists, critics say.
"I'm skeptical that it will be able to show that there is 100 percent accuracy in facial recognition technology, especially in using facial recognition technology out of a crowd," said Melissa Ngo, director of the Identification and Surveillance Project at the Electronic Privacy Information Center. "We've seen any number of studies and examples when trying to use facial recognition technology itself has been completely flustered when the subject is not standing still or in the right light, looking right at the camera."
The Glasgow didn't begin as a study of facial recognition programs, Jenkins said. Instead, he and Burton were actually studying the way the human mind recognizes both familiar and unfamiliar faces in a series of photos.
"If you [show ] unfamiliar people two photographs, whether they show the same person or not, [they're] not reliable at all. The reason is two photos of the same face can be very, very different images. The face itself can change a good deal as we age, due to fluctuations in weight or health. Even more important, the ambient conditions that the photo is taken in can change," Jenkins said. "These kinds of things have a huge impact on the image the pattern of light and dark across the page, but they don't tell us anything about who the person is."
With familiar faces of friends and families and celebrities, however, Jenkins and Burton found that people easily recognized the faces, regardless of age, angle or light in which the photo was taken. Faces that were "averaged," or blended from several images into one, were even more recognizable and yielded even better results with human subjects.
"It was this model that made us think, if we've captured something about the face-earning process, maybe we can use this same process to boost performance in automatic face recognition," he said. "Basically, we were trying to build a model of familiarity into the machine."
In the study, researchers used the web site MyHeritage.com to test their theory. The Web site uses face recognition software called FaceVACS, which is currently being tested at the Sydney Airport. The database at MyHeritage consists of 30,000 photos of celebrities.
Burton and Jenkins say when they submitted photos of 25 male celebrities, the software only recognized half of them. When they submitted their "averaged" images, which were blended by their new program from 20 images, the site recognized 100 percent of them.
"We think there is a kind of a virtue to this approach," Jenkins said. "It's a technique; it's a not a device. We're just saying, keep the machines. If you feed them different input you can vastly improve performance."
But critics like Ngo point out that even if the technology was perfect, there are social and legal issues at stake, such as the public surveillance systems that would likely be implemented to use facial recognition technology, arguing that such public systems are an invasion of privacy and not a deterrent to violent crime.
Ngo also points out potential civil rights abuses if someone is arrested based on a misinformation in a database.
"Another problem is the databases that contain all this information. They have incorrect information and they have outdated information," she said. "Now there's technology that allows individual law enforcement officers or commercial entities to 'identify' innocent people as a 'criminals.' ... Just because it is new, does not mean that it should be used. We need to figure out what standards that we're going to use or whether or not the technology meets its goals."
Despite critics of facial recognition, Jenkins is hopeful that this technique could be helpful, not harmful.
"I would hope that it would solve some problems," he said. "The cost of misidentification can be so high that I think we have to ask ourselves, given that these systems are starting to be deployed anyway, do we want to deploy a system that's reliable or a system that's unreliable?"