aces in a crowd," wrote Ezra Pound.

Take a look around the city — every sidewalk, every restaurant, every subway station is filled with strangers. We see so many new faces each day they become a blur, nameless, forgotten.

Unless, that is, one happens to have access to a closed circuit television camera network and a computerized facial recognition program.

In that case, Pound's anonymous New York faces, the "petals on a wet, black bough," become clusters of data points in an algorithm, sorted, analyzed and ultimately, identified.

Powerful facial recognition programs have recently attained a level of sophistication and reliability that allows them to be commercially viable. For example, Visionics, a New Jersey technology company, has introduced a facial recognition software package called FaceIt.

The software grabbed headlines when it was put to work in a Newham, a London neighborhood, in an attempt to rid the area of recidivist criminals. The software can scan up to 10 faces in a camera field and match them against filed images at a rate of 60 million records a minute.

"These are people that have been arrested," said Frances Zelazny, director of corporate communications for Visionics. "About 100 criminals were responsible for 85 percent of crime in the area. They set up 250 closed circuit cameras linked up to a control room…and very quickly criminals decided they would go elsewhere. We will be seeing replicas of Newham throughout the U.K."

The Visionics product comes out of research done at the computational neuroscience laboratory at Rockefeller University in Manhattan, where scientists studied the complexities of modelling vision by computer.

An example of the features identified by local pattern analysis. PHOTO: Visionics Corp.

Researchers Joseph Atick, Paul Griffin and Norman Redlich became convinced that the best way to design a system that could recognize facial characteristics was through what is called local pattern analysis. The method depends on an algorithm that identifies certain unchanging characteristics called nodal points-the distance from the eye socket to the bridge of the nose, for example. The FaceIt program can identify 80 nodal points, but only requires between 14 and 22 to make a positive identification.

Local pattern analysis is one of two primary tactics used by researchers in the field today. The other, called the Eigenfaces or principal components analysis, has also had success. The Eigenfaces model is akin to analysis of a color spectrum.

Just as a spectrum progresses from violet through blue, green, yellow and red, so too the image of any particular face can be broken down into a set of composite parts. If the analysis is applied to a large database, the result is a set of master faces, or Eigenfaces. These faces form a sort of template against which any individual can be identified.

Computer groups at both Rockefeller and the Massachuesetts Institute of Technology have successfully demonstrated the versatility of their methods in formal trials sponsored by the U.S. Army Research Laboratory. As a result, commercial applications may pick up speed here in the U.S. Soon, the person watching the camera may be a computer.

Back to top