Picking Out Faces In The Crowd

Cameras may be everywhere, but recognizing faces in the images is difficult for both people and machines alike. Test subjects do well at recognizing the faces of people they know in photographs. (Tony Blair is as easy to spot in a profile taken in a dark room as in a full frontal shot under the sun's glare.) But they do poorly at matching images of an unfamiliar person taken in varying light conditions, resolution, quality and camera angle. Computers have the same problem. The face-recognition system used by the Australian Customs Service, for instance, scored a poor 54 percent hit rate in a test by University of Glasgow psychologist Rob Jenkins, which he published recently in the journal Science. However, Jenkins was able to improve that score markedly by feeding the system a digital blend of several different images, rather than any particular one. Blended images pushed up the hit rate to an astonishing 100 percent. When Jenkins repeated the test using only those people whose images had stumped the system in the first experiment—a zero hit rate—it recognized 80 percent correctly. Jenkins believes that averaging helps "stabilize the face image," and could lead to more-accurate security systems. Big Brother just got a whole lot smarter.