Accuracy of facial recognition experts confirmed by researchers
Researchers have determined that trained forensic examiners are much more accurate identifying faces than non-experts and computers, providing the scientific basis for forensic practices, and pointing to possible methods of improving automated facial recognition systems.
In the experiment, facial images which were difficult for computer algorithms to distinguish between were shown to experts and non-experts, who were asked to identify matches.
Jonathon Phillips, a facial recognition researcher from the National Institute of Standards and Technology (NIST) says it is “the first strong evidence that facial forensic examiners are better at face recognition than the rest of us.”
Phillips conducted the research along with colleagues from the University of New South Wales and the University of
Texas at Dallas, addressing a call made in a 2009 report by the National Academy of Sciences to provide scientific support for forensic practices. To do that the research team tested and compared specially trained examiners, people working in the facial recognition field that are not trained examiners, and students. Images shown to them were upside down, included backgrounds, or were otherwise challenging for computer identification systems.
The researchers’ results are reported in the Proceeding of the Royal Society of London.
“We developed this test to find out if examiners recognize faces differently than untrained people,” Phillips explains. “The results indicate that they do, and we think that their training may teach them to use additional cues. We want to learn precisely how the examiners tell the difference.”
Using the same model of practice as forensic labs of engaging several examiners for face comparison decisions, the facial forensic examiners achieved a mark of 0.997, or nearly perfect recognition. This was despite time limits and no access to the specialized tools examiners have available in laboratories. Future experiments will study the accuracy of examiners in those familiar circumstances, Phillips says.
Once the criteria and methods used by trained examiners are identified and understood, they could presumably be incorporated into algorithms used by digital facial recognition systems.