| Electrical and computer engineering student Lingmin Meng (ENG'01) is engaged in a sophisticated identification game — he is developing mathematical algorithms that will allow computers to differentiate faces from the background in black-and-white images. Working in the laboratory of ENG Professor Truong Nguyen, Meng is developing a technology that will form the basis for a broad range of new applications, from personal identification and access control to enhanced low-bandwidth communication for video phone and teleconferencing.
Face detection, the ability to locate a face and ascertain its size, is different from identifying a face by matching it to one in a database. "Detection is the more fundamental and difficult problem," says Meng. "There is a huge range of variation to take into account — differences in illumination, direction, expression, and features such as hair, hats, and eyeglasses." Too, the size of the face relative to the background may vary greatly depending on how far the camera was from the subject. And background features such as buildings, plants, furniture, and the like can complicate the picture.
Meng's system is based on a three-tiered approach. First the algorithms compensate for differences in lighting by normalizing the contrast. Second, they identify the particular patterns of black and white pixels common to each facial feature (eyes, nose, and mouth), despite variations from face to face. Finally, the algorithms correct the image for distortions due to a tilt or rotation of the head as well as variations in size relative to the background.
The technique has thus far performed well in tests on standard sets of data. Meng is working to refine the process further, creating a core technology that may someday allow you to walk up to an ATM and, after a brief scan, be greeted by name.
|
