The new "Mugspot" software module developed at the University of Southern California automatically analyzes video images, looking for passers-by. When it finds them, it picks out the heads in the images and then tracks the heads for as long as they remain in the camera's field.

Throughout this tracking process, the software is watching for the best possible view of the subject's face -- the one that shows him or her looking most directly at the camera. It selects the best view presented and passes it on to the main face-recognition program.

This face-recognition software, developed at USC and the University of Bochum, Germany, and now in commercial use for clients such as Germany's Deutsche Bank, is robust enough to make identifications from less-than-perfect face views. It can also often see through such impediments to identification as mustaches, beards, changed hair styles and glasses -- even sunglasses.

"Until now," says Christoph von der Malsburg, the computer scientist and brain theorist who developed the system, "face- recognition software has needed to have the raw material, the images of faces, given to it in a highly structured form: a clear still photograph of a subject looking right at the camera.

"Our existing system is able to make identifications even with substandard images. With the addition of the 'Mugspot' video processing system, which expands its ability to capture images, I think it will prove useful in many real-life situations, particularly in law enforcement," says Dr. von der Malsburg, a professor in the USC School of Engineering's department of computer science.

Cameras mounted in airports and bus stations, or aimed at oncoming cars at traffic intersections, might continuously watch for known fugitives, von der Malsburg says. Bank surveillance cameras could identify persons seen at previous bank robberies.

The Mugspot system can scan eight video frames per second in real time, and takes about 13 seconds to select the best view, process it for identification, compare it to the several hundred faces in its memory and decide whether it has found a match.

The three research associates who developed Mugspot with von der Malsburg -- USC graduate student Egor Elagin, postdoctoral researcher Hartmut Neven and Bochum University visiting graduate student Johannes Steffens -- believe further refinement of the system can shorten that time by half.

Mugspot is only the latest improvement in the USC/Bochum face recognition software, developed with funds from the Army Research Laboratory (ARL) and marketed commercially in Europe under the trade-name ZN-Face.

In tabulations released Aug. 19 by the ARL, the USC/Bochum system outperformed competitors from laboratories across the country including MIT, the University of Maryland, Rutgers, Michigan State University and three systems developed by the Army itself.

The labs were ranked according to their performance on each of 12 separate runs, from 1 (best) to 12 (worst). The USC/Bochum system recorded six first, three seconds and two thirds, for a total of 18. The next best score was 24 -- a tie between MIT and the University of Maryland.

The USC/Bochum system also shone in tests conducted under substandard lighting conditions: It lost only a small fraction of its accuracy, while competitors showed drastic falloffs in less- than-brilliant illumination.

The USC/Bochum system uses an unusual approach that mimics the technique scientists believe the brain uses to recognize images. Von der Malsburg, whose principal research interests lie in the investigation of living brains, in fact carried out much of the original research on the system as part of an attempt to understand human face recognition. His research led to creating a computer model of the way the brain's visual cortex processes information.

Note: If no author is given, the source is cited instead.

• Artificial Intelligence
• Photography
• Software

• Computer vision
• Morphing
• Application software
• Webcam
• Videoconferencing
• Computer animation