Face Recognition System


Author: Ed Cohen

“Halt, who goes there?” sentries traditionally bark. But what if they didn’t have to rely on a password or coded response? What if a computer could tell friend from potential foe on sight?

Such is the dream of today’s military and security planners. Spurred by the terrorist attacks of September 11, computer systems are being developed that attempt to identify people from afar by the features of their faces or hands, the shape of their heads, even the way they walk.

The hope is that computerized monitoring could spot known terrorists in airports and other public places before they can do any damage. But the reliability of fledgling systems is far from perfect. In an interview earlier this year the head of a company that makes one face-recognition system acknowledged as much when he said his clients install the system more as a deterrent than because they believe the technology will actually catch criminals.

To evaluate the performance of vision-recognition systems and possibly develop new theories and systems, a pair computer scientists at Notre Dame has begun assembling what they say will be the largest inventory anywhere of digitized faces.

Every week 60 people — mostly undergraduate students — troop to Kevin W. Bowyer and Patrick J. Flynn’s new computer vision laboratory in the Fitzpatrick Hall of Engineering. In the lab and out in the hall the subjects have their picture taken 13 times with conventional, 3-D and infrared digital cameras under a variety of lighting conditions. To keep them coming back, volunteers receive a $10 gift certificate to the campus bookstore on every visit after their first. The pictures (minus names) will be made available to researchers worldwide later this year through the National Institute of Standards and Technology.

“The idea is to describe mathematically a person’s appearance,” says Bowyer, chair of the Department of Computer Science and Engineering and Schubmehl-Prein Professor of Computer Science.

In face-recognition systems, images captured by digital camera are simplified down to the salient features and turned into strings of numbers. As the computer captures new images and simplifies them it can compare them against the strings in inventory.

Bowyer and Flynn’s work is supported by grants from the Defense Department’s Defense Advanced Research Projects Agency, which became keenly interested in remote recognition systems after the bombings of the U.S. embassies in Africa. Back then the goal was to develop a system that could spot strangers from 50 or 100 yards away — before they could get close enough to explode a bomb. After September 11, the emphasis was broadened to include identifying known trouble makers at closer range in places like airports and border crossings.

The researchers say fingerprints remain the most reliable “biometric” or body measurement for use in telling people apart. But fingerprint-reading is too slow to use in monitoring crowded public places. Plus it relies on the cooperation of those being watched.

On the other hand, face-recognition faces seemingly insurmountable challenges. For one, a person’s face changes from week to week or even day to day based on facial hair, weight gain and other variables. Flynn, an associate professor, says there may be a greater difference between a person at different times than between two different people. And what about disguises?

Bowyer and Flynn think an effective recognition system will have to use a combination of technologies, possibly including infrared imaging, which looks beneath the surface to show heat coming from skin.

Say Bowyer, “A person could look at an infrared image and think, that face looks really cold. Maybe it has a lot of makeup on it.”

Ed Cohen is an associate editor of this magazine.

The magazine welcomes comments, but we do ask that they be on topic and civil. Read our full comment policy.