Deep Texture-Depth-Based Attention for Face Recognition on IoT Devices

Abstract

Traditional face recognition systems use RGB images as input for feature extraction and classification. However, conventional methods based on color images experience non-trivial accuracy drop under several challenging conditions like occlusion, pose variation and facial expression changes. With the gradually decreasing cost of smart sensors, RGB-Depth(D) images captured using low-cost sensors are used to provide complementary features to RGB images. Both the extracted Local Binary Pattern (LBP) features and depth map contain additional discriminative information that can guide the face recognition model to focus on the important parts of the input image. In this paper, we propose a novel end-to-end network that combines both texture and depth features for automatic attention-based face recognition. The experiment results demonstrate that the proposed method has improved recognition accuracy under diverse variations. Our proposed face recognition model has been implemented on the NVIDIA Jetson Nano device to evaluate its performance with compact feature extractors used on different branches of the model. The results show that our method can improve the FPS of face recognition on an edge-coming device from 1.6 to 3.8 with <1% accuracy degradation.