SMILE鈥攜ou鈥檙e on CCTV. But you鈥檇 better not grin if you鈥檙e trying to go undetected. It turns out that face-recognition systems have a much better chance of matching a suspect to an image if the pictures in their databases are of smiling faces rather than blank looks.
Face-recognition systems are now pretty good at matching a person to a stored image, but only when the database is relatively small, says Yaser Yacoob, who researches computer vision at the University of Maryland. 鈥淎s you start having thousands and even millions of faces in a database, they are going to look more and more alike,鈥� he says. So systems that compare large databases of faces with even larger numbers of faces in a crowd are likely to make lots of mistakes: more mugshots means more false positives.
Yacoob and his colleague Larry Davis realised that facial expressions might help software tell people apart better. Even people who look very similar reveal different features when they smile, says Yacoob, because it reveals more about bone structure and muscle structure. And changes to the surface of the skin produces different shading.
Advertisement
To test their theory, they used an image-processing technique called Principal Component Analysis, which is widely used for face recognition. It reduces an image to its most important features and stores these as a digital facial 鈥渟ignature鈥�.
Yacoob and Davis used the technique to compare the signatures of 60 smiling faces with the signatures of the same faces wearing neutral expressions. The program found a much greater degree of variation in the smiling faces than the neutral ones. That means there should be fewer very similar signatures and fewer false positives when searching a large database of faces, Yacoob says.
For example, a face-recognition system developed by Visionics of New Jersey was recently used to reduce fraud in a Mexican election. It had about three million faces in its database, so its failure rate of 3 per cent meant that 90,000 of the people could have been false matched. Joseph Atick, who runs the company, sees it differently: 鈥淭he way we look at it is that 97 per cent of the time we are able to eliminate fraud and only 3 per cent slip through the system.鈥�
Atick acknowledges that there are potential problems with building bigger and bigger image databases, but he isn鈥檛 convinced that a database of smiling faces would be useful. 鈥淭he problem is that recognition won鈥檛 happen when people don鈥檛 smile,鈥� he says. The system would somehow have to compensate for differences in expression.
But Yacoob thinks smiling faces shouldn鈥檛 be too hard to pick out because people often smile without thinking about it. The study also revealed that angry, grimacing faces are even more distinctive than smiling ones, but grimaces tend not to occur as naturally as smiles.