Embedded resource-constrained face recognition system for access control applications

Abstract

Propelled by the advancement of computer vision, facial recognition access control systems have been gaining ground. Highlighted by its accuracy and practicality, these systems have permeated personal mobile devices and are now integral to international border control security systems. When based on deep learning algorithms, improving its reliability usually implies increasing its complexity, adding more parameters and layers, and performing training with a larger and more sophisticated dataset. With the complexity augmentation, those systems inference is limited to powerful devices, leaving no space for resource-constrained applications. Therefore, this work pretends to develop and optimize a facial recognition system based on deep learning in a resource-constrained environment, analyzing the project adaptations that should be made to pursue a reliable and performing system with minimal resource usage. To perform that, lightweight deep learning models are used as a baseline, together with post-training quantization to reduce model footprint and increase the system performance. As a result, it was possible to have an entire face recognition system embedded in a Cortex-M7 device, with face detection, face anti-spoofing and face recognition blocks achieving a total inference time of 2,4 seconds.