CRIMINALS who go under the knife in an effort to evade capture might want to consider an alternative disguise, thanks to a new technique for matching faces before and after plastic surgery.
Typical facial-recognition software can be thrown off by even minor changes in the lighting and position of an unaltered face. Post-surgical matching is even harder for obvious reasons, says Kevin Bowyer, a computer scientist at the University of Notre Dame, Indiana, whose team developed the new system: "If someone has plastic surgery, they're trying to change the appearance of one or more parts of their face." As a result, existing software's success rate can be cut in half when trying to match before and after photos gathered from plastic surgery websites.
Bowyer's colleague, Gaurav Aggarwal, realised that matching individual facial features rather than whole faces could be more successful.
Aggarwal was inspired by a facial-recognition technique called sparse representation, which matches an image of a face by comparing it with combinations of individual features from faces already recorded in a database. If the closest matching combination turns out to be made up of features mostly drawn from one person in the database, it is a good bet to say the target image is also of that person. But if the best match combines features pulled from images of many different people then the system has failed to identify the new face.
However, to function properly sparse representation requires multiple images of each person in the database, so it does not work with pairs of before and after surgery pictures alone. The new system does. It uses two databases: a general one full of random faces, and another containing all of the "before" pictures - akin to police mugshots. When a target "after" picture is analysed, a composite picture as similar as possible is created from the features of people in the general database. All of the "before" pictures go through the same process. If the composite picture created using the "after" picture matches closely with any of the composite pictures derived from the "before" pictures, the two are declared a match.
The team found that while surgery changes the appearance of a face, many individual features stay the same, and matching based on the nose or eyes alone was actually more accurate than some existing whole-face techniques. Combining the matches of all facial features gave the team a 78 per cent success rate when comparing pre- and post-surgical photos. They presented their work this week at the Workshop on the Applications of Computer Vision in Breckenridge, Colorado.
"They're on the right track," says Christopher Solomon of VisionMetric, a company based at the University of Kent, UK, which provides facial ID software to police. He says the new approach could help police uncover disguised criminals but is unlikely to ever be totally accurate.
No comments:
Post a Comment