New method dramatically increases accuracy of facial recognition systems

The need to achieve higher accuracy at airport checkpoints and access- controlled sensitive areas lead security managers to look beyond fingerprints to other biometric technologies such as facial recognition. Trouble is, if airports used today’s face-recognition technology to identify passengers, they would often be wrong. The solution: University of Glasgow researchers have found a way to improve the technique by replacing the standard photo on a person’s ID with an image generated by combining several shots of the individual. The procedure dramatically boosted the technique’s accuracy. Even human beings sometimes have difficulty recognizing people from their photographs, especially if the focus is blurry or the lighting poor. As a face becomes more familiar, the brain improves at matching it to a photo. To explain this phenomenon, psychologists Rob Jenkins and A. Mike Burton came up with a model of how the mind constructs an image of a face from repeat encounters, distilling the essence of its features into a reliable mental representation. The researchers wondered whether applying the model to a face-recognition system would improve its performance.

ScienceNOW’s Yudhijit Bhattacharjee writes that this is exactly what they found when they tested the idea using a system that a few airports are trying out for small-scale recognition tasks, such as identifying airline crew members. The system is also used by a Web site containing a database of celebrity pictures at which visitors can upload personal photos to find out which celebrities they most resemble. To evaluate the baseline performance of the system, the researchers probed it with twenty different pictures of twenty-five famous men who were represented in the Web site’s database. On average, the system correctly identified the celebrity (by returning a photo of the same celebrity from its own database) 54 percent of the time. Using a computer program, the researchers then produced an “average” image for each of the twenty-five celebrities by merging each person’s set of twenty pictures, which had been taken over several decades under various light conditions. When they fed the averages into the system, it recognized the faces with 100 percent accuracy. The researchers then put the technique to a more difficult test: They constructed the average using only those images of a celebrity that the system had failed to recognize during the baseline performance test. This new average image was recognized correctly 80 percent of the time. The researchers report their results in this week’s issue of Science (see reference below).

How does it work? “Different photos of the same person can look very different depending on the ambient conditions at the time the picture was taken, the facial expression, and a number of other factors,” Jenkins explains. “By averaging these pictures, we’re essentially blocking out all that extraneous information.” The advantage of the technique is that it would be easy to implement; just swap the standard photos on passports and driver’s licenses with an average image. That way, when the airport camera snaps a picture of a passenger, a computer will be able to check whether that photo matches the ID more accurately. Face-recognition experts say the technique is worth exploring but needs to be tried on larger data sets. “This is too small a test set to make the claim that 100% accuracy has now been achieved,” says Anil Jain, a computer science professor at Michigan State University in East Lansing.

-read more in R. Jenkins and A. M. Burton, “100% Accuracy in Automatic Face Recognition,” Science 319, no. 5862 (25 January 2008) (DOI: 10.1126/science.1149656): 435 (sub. req.)