IMAGE RECOGNITION: Vision system finds faces in a crowd

A US-government-run competition between face-recognition systems was won recently by software called FaceIt, published by Visionics Corporation (Metuchen, NJ). The competition was part of FERET—a program funded by the US Defense Advanced Research Projects Agency that tested vision systems against a large database of faces. A system capable of recognizing faces as faces, in addition to identifying specific faces, would be useful for security applications and for visual information management, such as locating all the pictures of a specific person in a database.

Run by the US Army Research Laboratory (Ft. Belvoir, VA), FERET systematically tested various automated face-recognition systems using a single database of thousands of faces that includes a wide range of ages, races, and poses.

At the US Biometrics Consortium meeting (San Jose, CA) earlier this year, the Army Research Laboratory announced that FaceIt placed first in all four testing categories, namely basic recognition (matching two frontal face images of the same person), minimizing the number of wrong matches and correct matches not detected, matching images of people taken at different ages and in different places, and matching people in images taken in different poses.

Perhaps surprisingly, the program is commercially available and runs on Pentium computers under Windows 95 or Windows NT. FaceIt does not require any special hardware and uses widely available technology, which means it is notably less expensive than many of the other systems tested.

Recognition strategies

Face-recognition systems typically use a combination of strategies. According to Joseph Atick, Visionics president and a professor at the Computer Neuroscience Laboratory at Rockefeller University (New York, NY), the program uses a mathematical construction called local feature analysis (LFA) "to automatically derive a local topographic representation for any class of objects—such as human faces—from an ensemble of examples." The method is similar to an earlier method, principal component analysis, or eigenfaces, but LFA overcomes some of that method`s problems. Eigenfaces, says Atick, are global face representations that are sensitive both to deformations in the face and changes in pose and lighting.

Instead of looking at the whole face (a global representation) as the eigenfaces method does, LFA considers individual features, such as nose, mouth, cheek bones, or jaw line. Atick is emphatic about the importance of local, rather than global, representations. He suggests thinking of a low-dimensional representation as if it were a model made of LEGO building blocks: a face built of a few stereotyped components. In contrast, a global representation might look at each of thousands of pixels in a face image. "Pixels are bad," Atick says, "because you need too many of them. We figured out a way to build faces in terms of local blocks."

FaceIt combines LFA with the "eigenheads" method, which is also original to the Visionics researchers. Eigenheads is a low-dimensional representation of the three-dimensional (3-D) human head, obtained from shading information in the two-dimensional image. The eigenhead representation is independent of lighting and the 3-D pose, say the researchers. The program uses eigenheads to compensate for lighting and pose variation and then feeds the normalized face to the LFA code, which constructs a unique "faceprint." The faceprint can be matched to a database in real time and can obtain images from either live video or from static images.

Applications

The company has developed several products using the basic program. A personal-computer version protects computers or computer networks by allowing access only to users with known faces. Comparison of a prospective user`s face with a database library takes less than 1 s (see figure). Another product is a database-search program that compares a photograph of a person with a database and returns the best 25 matches.

Law-enforcement applications will be discussed by Atick, Norman Redlich, and Paul Griffin, the three Rockefeller researchers who started Visionics, at Photonics East (Boston, MA) in November (paper 2932B-22).