"So, this is the original source image..." Allen Yang, Research Scientist at UC Berkeley, points to a face on a large computer monitor.
The conventional way to recognize a face is to record distance between the eyes, width of the mouth, shape of the nose. But new research shows that recording random points is faster in crowds and more effective with disguises. That research by UC Berkeley research scientist Allen Yang exploits a new algorithm called "sparse representation".
Think of a face in a crowd as a flavor in a mixture of juices. It's very difficult to taste any one of the flavors. So Yang's new approach picks out faces in a crowd like the ingredients on a label. Each flavor, or sparse representation, appears as a spike on a graph.
The algorithm doesn't need a high resolution picture. In fact, you can throw out almost all the detail, all the color, and be left with only 200 pixels. That is all the computer needs to draw the highlights out of an image -- the sparse representation -- and match an image that it knows in its database."
"This actually shows that our algorithm has the potential to outperform some of the human ability to recognize faces," says Yang.
Shankar Sastry, Dean of Engineering at Cal, was also involved in the research. He adds, "We'd actually like to be able to outperform human analysts who look at images perhaps with the kinds of obscuration that we expect bad guys to use." For example, a mask. The new method breaks an image into 8 sparse representations, so it can isolate the mask, then reconstruct the complete face.
The next step for this group is to link your face to your voice and to the way you walk -- multimodal data. Research subjects are even now wearing the sensors to build a database of gates. So, even if I can't see your face, I could pick you out of a crowd by the way you walk. Cal engineers recognize the privacy implications and anticipate that public agencies will need to establish some safeguards.
------- Links -------