recognition

Researchers have developed a new face recognition program that uses a new mathematical model that increases the accuracy of recognition even in cases of disguise, varying expressions, or poor image quality. The sparse representation algorithm can match an image to one in a database regardless of major facial occlusions or image corruption.

U. ILLINOIS (US)—An engineering team has developed a face recognition system that is remarkably accurate in realistic situations.

Unlike existing face recognition programs that try to find “optimal” facial features, the new program uses sparse representation. One of the program’s developers, Yi Ma, an associate professor at the University of Illinois, contends that the choice of features is less important than the number of features used.

“Face recognition is not new, but new mathematical models have allowed researchers to identify faces so occluded that it was previously thought impossible,” says Ma.

People can learn upwards of tens of thousands of different human faces during their lifetime. Various real-world situations such as lighting, background, pose, expression, and occlusion may complicate human recognition, but are incredibly difficult problems for traditional face recognition algorithms to conquer.

Ma’s sparse representation algorithm randomly selects pixels from all over the face, increasing the accuracy of recognition even in cases of disguise, varying expressions, or poor image quality.

The algorithm also increases accuracy by ignoring all but the most compelling match from one subject.

Experiments using sparse representation support the approach. In an experiment that uses two established databases of faces, the Yale B and the AR, the new face recognition method is remarkably accurate. Applying this approach to the Yale B database shows 98.3% accuracy using mouth-region images. The AR database shows 97.5% accuracy on face images with a sunglasses disguise and 93.5% accuracy with a scarf disguise.

The technology is jointly owned by the University of Illinois and the University of California, Berkeley, and could have applications for personal and corporate use.

“The computer can identify images that the human eye can’t,” says Ma, who sees a future where people can capture someone’s face with their camera phone, upload the image to a web-based service, and have a match sent to them seconds later.

University of Illinois news: http://news.illinois.edu/