Academic researchers develop method for improving facial recognition accuracy with poor image quality
Research by psychologists from the UK and Australia suggests that the accuracy of facial recognition using images from poor quality CCTV footage can be improved by combining them with images from other footage in a computer-enhanced “face averaging” composite, Phys.org reports.
The academics, from the universities of Lincoln, York, and New South Wales compared the effectiveness of humans and automated facial recognition systems at identifying people from high quality images, pixelated images, and face averages, and found that both humans and computers were more accurate with combined poor images than with individual poor images. Further, computer systems were more accurate with combined high-quality images than individual ones, in some cases reaching 100 percent accuracy.
“We know that not all CCTV systems have the luxury of high quality cameras, meaning that face identifications are often being made from poor quality images. We have shown that there is a relatively quick and easy way to improve pixelated images of someone’s face,” says Dr. Kay Ritchie of the University of Lincoln’s School of Psychology. “We also know anecdotally that there are lots of different techniques that people can use as investigative tools to improve low-quality images, such as manipulating brightness. Our standardised face averaging method could help in suspect identification from low-quality CCTV footage where images from multiple different cameras are available, for example, from tracking a suspect along a particular route.”
Participants in the study were asked to compare a high-quality image with either a low-quality image or one created with the averaging method, and were significantly more accurate when viewing combined images. A smartphone application and a commercial facial recognition system the researchers say is widely used in forensic settings both likewise showed higher levels of accuracy with composite images.
Google researchers recently developed a method for avoiding the traditional trade-off of slowness for accuracy of facial recognition applications constrained by hardware capacity.