Picture This

Rising accuracy rate makes face recognition technology a viable security tool.

Eighty percent is a decent success rate in most cases. But for a security system meant to deny entry to terrorists and other criminals, while letting others pass through quickly, it's not good enough.

When the National Institute of Standards and Technology conducted a test of face recognition products in 2002, it reported a 20 percent error rate-the number of innocent people falsely identified as security risks-for the three technologies evaluated. That was a marked improvement over the 54 percent error rate in a 1997 government test and a nearly 80 percent error rate in 1993.

"One out of five people having a mistake-that is a carnival trick, not a security solution," says Jeremy Grant, a senior vice president and identity solutions analyst at Stanford Washington Research Group in Washington.

This spotty history does much to explain the excitement over the 2006 face recognition product test, the results of which NIST released in late March. The error rate for some systems is now down to 1 percent. Grant and others say the orders-of-magnitude improvement marks the arrival of face recognition technology as a legitimate security tool for federal agencies and private entities.

In a test involving still images, Google subsidiary Neven Vision of Santa Monica, Calif., was the most accurate system, registering a 1 percent false rejection rate. When NIST used 3-D images, the face recognition technology from Viisage, a subsidiary of L-1 Identity Solutions Inc. of Stamford, Conn., won with false rejection rates ranging from 0.5 percent to 3.1 percent.

What really drives home the viability of these tools, according to the NIST test, is that automated face recognition technologies are capable of equaling or surpassing the performance of an average person perusing mug shots. Undergraduate students at The University of Texas at Dallas were shown 80 pairs of photographs for two seconds at a time and asked whether the two faces were the same person. They also had to rate their confidence in their answers. Six out of seven automated systems outperformed the students.

Improvements in face recognition technology are due to several factors. Advances in algorithms played a large part, increasing the technology's performance by a factor of four to six, according to NIST. But the dramatic upswing, the study noted, also was a result of higher resolution images and more consistent lighting. So as cameras improve, face recognition technology gets better too.

While the 2006 test showed a marked improvement in the technologies' ability to match faces across changes in lighting-systems made by Neven Vision, Viisage, the Samsung Advanced Institute of Technology and the German firm Cognitec Systems all fared better than the top-rated products in a January 2005 test-the accuracy of face recognition systems drops dramatically as the illumination for the photographs becomes less standard, the NIST report said.

Grant says this means that face recognition is still a stretch for access control systems that use cameras to scan crowds and attempt to match faces against photographs in terrorist and criminal databases. Settings where photographs are taken with a standard illumination and background-driver's licenses, passports and other credentials-are a better fit.

NIST also studied iris scanners and found that the most accurate one, a combination of technology from French company Sagem Morpho and L-1 subsidiary Iridian Technologies, performed impressively, with an error rate of about 1 percent. But in the first head-to-head evaluation of multiple biometrics, NIST found that iris- and face-based recognition systems had "comparable accuracy." Grant says this is a blow to manufacturers of iris scanners that have touted their products as more accurate than face recognition.

Ultimately, Grant says, federal agencies, law enforcement authorities and others could opt to use systems with multiple biometric components. Checking fingerprints against an FBI database while also comparing mug shots against another database increases certainty about identity and decreases the chances of misidentifying a criminal as harmless.