Imagine never worrying about forgetting your bank card PIN number, as your ATM always recognizes your face. Or if the police could track down wanted criminals by spotting their likeness in a crowded train station.
That’s the vision that drives researcher M. Alex Vasilescu. Although a degree of facial recognition technology exists today, it remains largely inefficient and too easy to trick.
It’s a problem Vasilescu has devoted herself to solving. The University of Toronto PhD candidate is currently conducting research at New York University in the area of face recognition software.
The field of biometrics – technology that combines IT security and human physiology – is booming. Fingerprinting dominates the field today, but facial recognition is predicted to be a fast-growing market segment, projected to be worth US$800 million by 2008.
Vasilescu’s research is already having an impact – her research has attracted the attention and funding of the U.S. Department of Defense. The famed Massachusetts Institute of Technology (MIT) also recently named her one of the top 35 researchers under the name of 35 for her groundbreaking work.
While in Toronto recently speaking at a meeting of the Canadian Information Processing Society, Vasilescu explained her research. “Face recognition is a difficult problem for computers,” she said. “An image (of a face) is made up of multiple factors,” including the face geometry itself, the amount of light it’s exposed to and the shadows that creates, from what angle a face is viewed and facial expressions.
“Computers are much better at answering the question, ‘are those the same images?'” she said, versus recognizing a single face.
The problem then becomes, how do you extract what’s important from a given image of a face so that a computer may always make a match, regardless of lighting or perspective? One popular method is based on Principal Component Analysis (PCA), which uses common linear algebra to come up with statistical relations within an image.
But, says Vasilescu, this methods works best when only one factor, say lighting, is varied. Multiple variations introduce more problems.
Enter Vasilescu’s approach, known as TensorFaces. Adapted from the technology used to render life-like motion in computer animated films such as Toy Story, it too takes a face image and breaks it down into its constituent factors, this time using multilinear – or tensor – algebra, and builds a recognition system based only on the person’s identity. “It throws away issues related to lighting and viewing direction, but keeps issues related to people recognition,” Vasilescu said.
However, the true value of face recognition, from a security standpoint, is to be able to render faces on the sly, without the cooperation of the subject. Vasilescu and her team, using volunteers, were able to model how faces change in different circumstances, making it possible to predict what Osama bin Laden, for example, would look like seen from a random viewpoint.
In lab tests, face recognition rates for the TensorFaces approach has scored well. But Vasilescu stressed that her work is still much lab-based. As to whether the technology could truly be licensed today to spot fugitives in a crowd of thousands in a real-world setting, “that remains to be seen,” she said.
“They’d like to have it up and running in a year or two.” For those who think the notion is far-fetched, Vasilescu refers them to Walt Disney World in Orlando, Fla., where park pass patrons obtain cards with their handprint stored in their memory. They then use both the card and their hands to enter the site.
“I never would have thought Disney would be using biometrics, but there you go.”