It’s all in the pixels: How software matches faces
Look at any Facebook photo album, and there’s a good chance Facebook has already figured out who’s who in the pictures and has attached names to the faces, even if the original uploader did not.
It’s able to do that by using special computer programs that match faces in photos to faces its users have already identified.
Now, law enforcement agencies are using the same kind of technology to identify criminal suspects caught on camera.
There is a 60 percent chance that an average-quality surveillance photo could be matched to a photo of the same person in a mugshot or driver’s license picture database, according to Michigan State University computer science Professor Anil K. Jain.
With better, higher-resolution surveillance photos, accurate matching can approach 100 percent. “It all depends on the quality of the image you have acquired,” Jain said.
To test the technology, Jain added a high school yearbook photo of one of the Boston Marathon bombing suspects to a database of 1 million photos taken for Florida driver’s licenses.
Using off-the-shelf software sold to police departments, Jain compared a fuzzy FBI image of the suspect at the marathon to the huge database of photos. The software picked the yearbook photo as the No. 1 match.
Jain said the software uses statistical methods to identify and extract tiny groups of pixels that represent unique parts of significant facial features such as eyes, noses and mouths.
This reduces the number of pixels that need to be compared from over a million to just a few, making matches against millions of photos much easier.
Different software companies use different approaches, but the trick is to “find the right features that don’t change very much based on the conditions under which the [surveillance] photo was taken,” Jain said.
So good software will be able to match a head-on mugshot taken under controlled lighting to an odd-angled photo grabbed from fuzzy video.
Under real-world circumstances, the computer typically narrows a suspect’s photo to the 100 or so best matches, ranked by probable accuracy. “It’s much easier for a human being to look at 100 photos than 7 million,” Jain said.