The Baltimore Sun

Computers prove weak at faces

Identification: The job of creating a machine that can match a glimpse of a passing person with a database photo turns out to be more difficult than expected. 

September 27, 2004

By Dennis O'Brien

You'd think it would be easy for a computer to mimic something we do dozens of times every day: sizing up a face and putting a name to it.

It's not - a fact that's becoming all too apparent to those who were counting on the technology to protect our borders and airports.

Researchers say that identifying one another is a complex process, and after a decade of work, they've learned that designing a computer to reliably match photos with people is harder than it sounds.

"Our eyes and our brains do a lot of things that we take for granted. These are really hard problems," said Richard Mammone, a biometrics researcher at Rutgers University. "Mother Nature's an amazing thing, and we're just catching up."

Facial recognition systems are becoming a key tool for a security-conscious nation since the Sept. 11 terrorist attacks on the World Trade Center and Pentagon.

Systems that match employees' faces with photos as they arrive at work sites have been on the market for several years and are expected to account for $114 million of the $1.2 billion in biometric device sales this year, according to the New York-based International Biometric Group. Sales of facial recognition systems are expected to double next year, according to the group.

But those systems typically match close-up scans of live faces with a limited number of equally good photos - often taken with the same equipment. Reliable systems capable of identifying potential terrorists as they walk through airports and or other public facilities are still in their infancy.

"The low-hanging fruit has been plucked. Now it's a question of moving on from there," said James Wayman, a biometrics researcher at San Jose State University, in California.

The National Institute of Standards and Technology, the federal agency charged with evaluating biometric technologies, began promoting a "challenge" to researchers at a biometrics conference in Crystal City, Va., last week. It issued a public call to encourage scientists to come up with systems reliable enough to be used at airports and rail terminals.

But experts say it's difficult to design a program that can match an image of a real person by comparing it with thousands of photos that might be old, darkly illuminated or have other technical flaws. Factors such as facial hair and glasses make accurate matches in real time at relatively long ranges difficult.

"If the subject's face is directly shown or in profile, if the lighting conditions aren't good, all of these things are factors in the process," said Takeo Kanade, a facial recognition researcher and the director of the Robotics Institute at Carnegie Mellon University in Pittsburgh.

There are a variety of other biometric techniques. Fingerprint and hand-geometry scanners are used to give employees and customers access to work sites, schools, hospitals and amusement parks.

Some work better than facial scanners. NIST says tests earlier this year showed that when it comes to verifying a person's identity, fingerprint systems still work better than any of the 34 facial recognition systems that were tested in 2002.

But it's the facial scanner - with its ability to tap into vast collections of photos and do its job surreptitiously - that has the best potential for identifying fugitives.

"When you're dealing with a terrorist, you might not have a fingerprint or even a clear photograph. All you might have is a grainy image taken from a distance," said Amanda Goltz of the International Biometric Group, a research consulting firm.

Federal agencies are providing much of the impetus for facial recognition research and development. The State Department announced in August that it will begin requiring passports to include chips embedded with digital photographs that can be compared with images taken at border checkpoints. The system, designed to be consistent with those adopted by other nations, is supposed to be in place nationwide by December 2005.

In May the Department of Homeland Security awarded a $10 billion contract to Accenture LLP to come up with a system that will be able to conduct 440 million inspections a year at 300 points of entry across the United States.

The technology used to track visitors could include a combination of biometrics - fingerprint scanners, digital photos or some type of facial recognition system.

As part of its challenge, NIST is asking the nation's biometrics researchers to submit their facial recognition systems to it for testing to show whether they meet the NIST standard of a 98 percent accuracy rate for verification of photo identities. About 46 companies and universities have responded. There is no deadline for participation, but NIST plans to have test results available next fall.

There is no reward for meeting the standard - other than major bragging rights and a terrific marketing tool. "They'll have the pride of meeting this goal for the first time," said Jonathan Phillips, the NIST electronics engineer who discussed the challenge in a presentation at last week's Biometric Consortium Conference. The three-day event was sponsored by NIST and other government agencies that evaluate biometric devices.

In general, facial recognition systems analyze the position and spacing of facial features to generate a code that is matched against codes in a database of digital photos.

Mammone, one of several speakers at last week's conference, is working on a system that breaks up images of human faces into blocks or regions and turns them into pixels on a computer screen. Each point in a photograph - such as the chin, the nose and the shape of the cheekbones - is mapped and coded so that specific points can be matched with those in other photographs.

If the photo image is similar to one in the collection, the computer shows a high score. If there are no similarities, the score is lower.

Experts warn that even with government funding, a facial recognition system as effective as the human brain is probably years - if not decades - away.

"When will the computer be as effective as the human? I'd say it will happen, but not for tens of years. Maybe 10 to 30 years," Kanade said.