Comment

9 July 2009

Technology can see through a disguise

A disguise may work to fool the naked eye, but a new computerized face recognition system may be able to see through any kind of masking.

Every face has special features that define that person, yet faces can also be very similar, said Lin Huang, of Florida Atlantic University, in Boca Raton, Fla. That is where computerized face recognition comes in.

Face recognition software has been in development for quite some time. However, for biometric authentication at border crossings, for access to buildings, for automated banking, crime investigation, and other applications, it has not yet become a mainstream application. The main technical limitation is although the systems are accurate they require a lot of computer power.

Early face recognition systems simply marked major facial features (eyes, nose, and mouth) on a photograph and computed the distances from these features to a common reference point. In the 1970s, a more automated approach using a facial template extended this idea to map the individual face on to a global template. By the 1980s, an almost entirely statistical approach led to the first fully automated face recognition system.

In the late 1980s, researchers at Brown University developed the “eigenface method,” which was extended by a team at MIT in the early 1990s. Since then, approaches based on neural networks, dynamic link architectures, fisher linear discriminant model, hidden Markov models, and Gabor wavelets. Then a way to create a ghost-like image that would succumb to an even more powerful analysis could accurately identify the majority of differences between faces.

However, powerful techniques have so far required powerful computers. Now, Huang’s team applied a one-dimensional filter to the two-dimensional data from conventional analyses, such as the Gabor method. This allows them to reduce significantly the amount of computer power required without compromising accuracy.

The team tested the performance of their new algorithm on a standard database of 400 images of 40 subjects. Images are grey scale and just 92 x 112 pixels in size. They found their technique is not only faster and works with low resolution images, such as those produced by standard CCTV cameras, but also solves the variation problems caused by different light levels and shadows, viewing direction, pose, and facial expressions. It can even see through certain types of disguises such as facial hair and glasses.

For related information, go to www.isa.org/sensors.