1/13/2009 Kim Gudeman, Coordinated Science Lab
Beware thieves and other ne'er-do-wells: Those sunglasses and other disguises may soon no longer protect your identity if ECE Associate Professor Yi Ma and his group have anything to do with it.
Written by Kim Gudeman, Coordinated Science Lab
Beware thieves and other ne'er-do-wells: Those sunglasses and other disguises may soon no longer protect your identity if ECE Associate Professor Yi Ma and his group have anything to do with it.
Ma, who is a researcher in the Coordinated Science Lab, and his students have developed a facial-recognition algorithm that can identify an individual even if the image is corrupted or occluded. The algorithm works with 90 to 95 percent accuracy when the nose, eyes or mouth is obscured, either by disguise or a corrupted image.
“Face recognition is not new, but new mathematical models have allowed researchers to identify faces so occluded that it was previously thought impossible,” said Ma. “But the computer can identify images that the human eye can’t.”
Ma’s research takes facial-recognition technology leap years beyond current capabilities. Modern-day systems require high-resolution images, and have difficulty dealing with changes in lighting and expression. In addition, the systems can be tricked by adding disguises such as mustaches, sunglasses, or other occlusions.
The Illinois group’s system is more selective, choosing the smallest number of images from the database that can represent the new test image. The mathematical tools for selecting these images come from the theory of spare representation, which seeks solutions to linear equations in which almost all of the entries are zero.
Ma and his team, which includes researchers at the University of California, Berkeley, have known for a while that the algorithm works, but lacked a theoretical explanation for its good performance.
“It turns out that faces are more similar than models that people have studied before,” said John Wright, an ECE PhD student in Ma’s group. “That allows the computer to compare very highly correlated vectors and get an extremely accurate result.”
The technology has applications in security, setting the stage for better facial-recognition entry systems. Also, it could produce new methods of annotating video and still images. For example, it could allow you to find friends via a photo on social networking sites. Or you could search for your own face on the Internet to see if it has been used in an unauthorized way.
While the group has worked mostly with images, the mathematical model may also work with audio, helping to clean up background noise like a cough during a symphony, for example.
Ma and his team are currently integrating the math and the interface: “We’re working to bring the technology closer to real-world deployment.”