Facial recognition systems may be a staple of sci-fi thrillers and TV police procedurals, but in the real world they are still some way from becoming reliable enough to use either as a personal access identifier or as a security tool to pick out suspects in a crowd.
Facial recognition has been significantly affected by occlusions such as scarves and sunglasses, large pose variations, heavy makeup or poor image quality.
But this might be about to change. Research led by Artificial Intelligence specialist Professor Dacheng Tao in the Faculty of Engineering has led to significant advances on one of the key questions facing the technology: how to improve accuracy in real-time identification without needing an exponential rise in computing power and processing time.
Based in the School of Computer Science, Professor Tao attributes his success to two central advances: the invention of video data simulation for model training which helps systems to reduce the impact of image blur in discerning the key characteristics of faces, and the introduction of a multiview learning model named Trunk-branch Ensemble Convolutional Neural Network (TBE-CNN). TBE-CNN integrates information from the holistic image and multiple facial components to build up a more robust image representation that overcomes some of the challenges of pose variation and occlusions.
What is more remarkable is that these improvements have not come at the cost of massive increases in data-processing power. The new multiview system uses three different images to create a facial pattern but it uses less than 1.3 times the memory and time of a single-image system.
“We believe the performance can be pushed still further,” he says. Although facial recognition is at heart a pattern-matching problem, it presents unique challenges, but Professor Tao believes that none of them are insurmountable. In recent work, Professor Tao’s team have promoted the verification rate to 99.25% by a single deep model.
In a complete facial recognition system, facial detection is an indispensable step before facial recognition and this is where Professor Tao’s team has also achieved significant advances. Where standard cascade facial detection algorithms score around 60 per cent accuracy for every 1000 false positives on the Face detection Data set and Benchmark (FDDB) test, Professor Tao’s system is already achieving accuracy rates greater than 97 per cent. On the more challenging WIDER FACE benchmark test, where the standard cascade systems have average precision scores of less than 14 per cent, the new algorithm scores above 76 per cent. But Professor Tao is still not satisfied.
The age of mass surveillance is already with us, and Professor Tao believes that the more accurate a facial recognition system is, the less likely it is to mis-identify innocent people. The new system he has created marks a milestone on the path to a safer world.