Jain believes the Michigan State research, which is supported by the National Institute of Justice, is the first large-scale experiment that attempts to match forensic sketches to police photographs. The system was originally developed by a former Jain graduate student, Brendan Klare, now at Noblis. Current enhancements are being carried out by graduate student Scott Klum.
“To our knowledge, there is no sketch-to-photo system available on the market,” says Jain, adding that all commercial face recognition systems work well when matching photos-to-photos, but fail miserably in matching sketches-to-photos.
The novelty of the Michigan State approach is to find image features that are amenable to matching sketch-to-photo. In order to achieve this, the Michigan State team first trained the system using sets of sketches and photo mug shots obtained from police from previously solved cases that involved both a sketch and a mug shot.
“Our system learned the similarities between a sketch and a photo based on these solved cases,” comments Jain.
Results have been promising. Using a database of more than 10,000 mug shot photos, 45 percent of the time the system matched the correct person.
In fact, according to Jain, the system performs even better than state-of-the-art photo-to-photo matching systems in finding hits.
He now has a prototype of the system, called the FaceSketch ID System, which is currently being evaluated in the field by the Pinellas County, Mich., Sheriff’s Office and the Michigan State Police—the two agencies that gave the Michigan State researchers access to sketch-mug shot photo pairs of solved cases used to train the FaceSketch ID system.
And Birdwell believes Michigan State researchers may be on to something useful.
“It may become one more tool by which law enforcement can attempt to identify a person-of-interest,” she says. The system, she notes, doesn’t appear to be endangering any part of the investigative process by doing so, which is a positive step.