'Degraded' Synthetic Faces Could Help Improve Facial Image Recognition

Researchers from Michigan State University have devised a way for synthetic faces to take a break from the deepfakes scene and do some good in the world – by helping image recognition systems to become more accurate. The new controllable face synthesis module (CFSM) they've devised is capable of regenerating faces in the style of real-world video surveillance footage, rather than relying on the uniformly higher-quality images used in popular open source datasets of celebrities, which do not reflect all the faults and shortcomings of genuine CCTV systems, such as facial blur, low resolution, and sensor noise – factors that can affect recognition accuracy. CFSM is not intended specifically to authentically simulate head poses, expressions, or all the other usual traits that are the objective of deepfake systems, but rather to generate a range of alternative views in the style of the target recognition system, using style transfer. The system is designed to mimic the style domain of the target system, and to adapt its output according to the resolution and range of'eccentricities' therein. The use-case includes legacy systems that are not likely to be upgraded due to cost, but which can currently contribute little to the new generation of facial recognition technologies, due to poor quality of output that may once have been leading-edge.