Guess who? Helping machines recognize faces

Dr. Rob Jenkins, who carried out the research, with his current passport photograph (left) and his average face image (right). This photo is intended to illustrate the improved likeness brought about by the image averaging technique.

Unless you’re decked out in a really elaborate disguise, say as a Lord Voldemort or a Star Wars creature, you can generally recognize yourself in the mirror. It doesn’t matter what angle you’re facing or how the lighting is – you still know you’re you.

This image illustrates the central problem in face recognition: changes in lighting, pose, and facial expression can affect face images more than changes in identity. As a result, photos of the different people can be more similar than photos of the same person. The image above shows six photos of Person A, and six photos of Person B. But can you tell who is who? Answers: AABABA (1st row) / BBAABB (2nd row)

But, it’s not nearly so easy for computers.

A company might have a security system that includes a computer with photographs of all the employees stored in its “brain.” Then, when an employee wants to get in the door, he or she stands in front of a camera and the computer decides whether it’s “looking” at a face that matches the photos stored in its memory.

Unfortunately, people can look really different from one moment to another, depending on the lighting, angles, facial expression, etc. This doesn’t really bother the human brain, but computers have a hard time ignoring those differences and making face matches under these conditions.

Researchers have now found a way to improve automatic face-recognition programs.

Rob Jenkins and A. Mike Burton of the University of Glasgow in Glasgow, Scotland, discovered that merging several photos of the same person into one average image can significantly improve a computer’s accuracy when it matches a person’s face to an image in its memory.

Face-recognition programs are not particularly sophisticated right now, but people are interested in using this technology in the future as a tool for catching criminals.

The study appears in the 25 January issue of the journal Science.

###