When it comes to recognizing faces, the gap between computer and human error is closing rapidly, and in some cases facial-recognition technology is surpassing human ability — a significant advancement for agencies looking to use biometrics to improve security for their systems and facilities.
Facial-recognition technology has developed over the past 20 years, but variables such as lighting, positioning and facial expression make it harder to identify potential threats compared with fingerprint or iris recognition. Current systems are designed to work on fairly small, still facial images, like police mug shots. But researchers have made breakthroughs in the development of computer algorithms that recognize changes in position and lighting.
Better Than the Human Eye
“We’ve discovered that the best algorithms are better than human performance,” says Jonathon Phillips, test director at the National Institute of Standards and Technology. “Humans are good at recognizing people they know in a situation that you expect, but not nearly as good at recognizing unfamiliar faces. So we’d put pairs of faces up on the screen and ask humans if they think it’s the same person. They’d state their opinion, and then we’d ask the algorithms.”
The tests were conducted during NIST’s latest Facial Recognition Vendor Test (FRVT), the most recent in a series of federally funded test programs that began in 1993 with the Face Recognition Technology (FERET) program. The goal of these programs is to develop state-of-the-art technology and identify the most promising approaches to facial biometrics.
“This is an interesting time in the development of facial recognition,” says Theo Cushing, an analyst at the International Biometric Group (IBG). “There’s definitely a movement forward, and the technology is getting noticeably better. We have two divergent tracks of development: the 2D traditional facial recognition that’s being used by the departments of motor vehicles and for visa applications, and emerging facial biometrics like 3D and thermal imaging that picks up a face at a really long distance.”
Cushing is a consultant for the IBG’s Sensors, Surveillance and Biometric Technologies Center of Excellence, which is contracted by the Justice Department to support state and local law-enforcement programs.
“Investment and deployment peaked after Sept. 11,” Cushing says. “There was a lot of money flying around in products after that, and now we’re looking at what actually worked and what didn’t. The government is now defining real-world requirements.”
Testing has spurred major advancements in facial recognition, thanks in large part to the collection of vast data sets that have led to the development of new algorithms. Improvement of two orders of magnitude, or 100 times, has been documented between the launch of the FERET program and 2006 and one order of magnitude in the past four years, according to Phillips.
“In recent years we’ve tried to compare how well humans do different tasks versus algorithms,” explains psychologist Alice O’Toole, a researcher at the University of Texas at Dallas who worked with Phillips on FRVT. “We took the algorithms that did the worst; the most difficult task was matching pairs of images when the two images varied in illumination. This turns out to be very difficult for humans and algorithms. Of seven algorithms, we found three that performed better than humans and four that did worse, and that’s pretty good.”
Neither man nor machine is 100 percent correct, but when researchers combined all seven algorithms with human performance, the results were nearly perfect, O’Toole explains.
The More, the Merrier
Researchers are looking to multimodal biometrics technology that utilizes more than one biometric tool — finger, face or iris — to identify the enemy covertly under less-than-ideal circumstances, such as substandard lighting, long distances or changes in facial features.
According to an IBG report on the state of facial-recognition technology, these applications require fundamentally new approaches that “depart radically from approaches optimized for travel-document-based applications. New approaches may come from outside the established face-recognition market, from firms experienced in modeling faces for gaming or other nonbiometric applications.”
Cushing says, “You can think of facial biometrics by itself, but emerging biometric solutions are stitching together different modalities.”