Facial Recognition Software
After both the Brussels terror attack in 2016 and the poisoning of a former Soviet double agent in London 2018, law enforcement agencies made use of a mysterious tool in the investigative arsenal: the super-recognizer. An individual with the ability to recall thousands of faces after having seen them but once in passing, super-recognizers aided law enforcement agencies to identify suspects in both incidents. Often investigators have little choice but to appeal to the public for help identifying suspects in a terror attack, who appear on grainy security camera footage. Being able to recognize and match faces with an error rate of about three to four times lower than the average population makes these super-recognizers instrumental in the race against time to apprehend suspects as they flee.
While the human error component is minimized by relying on super-human facial recognition, the system is far from perfect. Both the success and speed of apprehending suspects seen only on low-quality film would greatly improve if recognition were automated, however computing capacity has remained a constraining factor in the deployment of such strategies. Currently, algorithms can recognize faces flawlessly – if they are from controlled, high-quality images, like passport photos. Once the faces are seen in poor lighting conditions, low resolution, and different poses, the automated systems fail. If law enforcement agencies are to have robust capabilities and move away from embedded super-recognizer networks reminiscent of Sherlock Holmes’ urchins, computational approaches must learn to navigate low-quality footage of suspects.
Yalda Mohsenzadeh and her research group in the Department of Computer Science at Western University are bringing together the worlds of computational neuroscience and AI to build facial recognition software by proposing an identity-preserving, super-resolution approach. “If we show a machine learning algorithm two photos of the same person, one high quality, the other low, it can learn how to turn grainy features into sharp ones,” she explains. A trained algorithm is able to predict what a grainy image of a random person taken from security footage would then look like in high definition. These “cleaned-up” images are then usable to cross-reference against a database.
The deep neural network-based algorithms developed by the Mohsenzadeh group are approximately 68% reliable on controlled, frontal, low resolution faces, meaning they are about three times as accurate as other state-of-the-art facial recognition endeavours. On real-world surveillance camera pictures, their method performs with 52% accuracy; twice as good as the state-of-the-art methods. International intelligence networks, like Five Eyes, of which Canada is part, are becoming increasingly connected. As global threats permeate national borders like never before, the ability for allied states to generate and share reliable information is vital.