Learning to Drop Expensive Layers for Fast Face Recognition

Recent years have seen many advances based on Deep Convolutional Neural Networks (DCNNs) in the tasks of face recognition, most of which are developed to pursue high recognition accuracy. In this paper, we propose a novel Fast FAce Recognizer (Fast-FAR), learning to improve the speed of DCNN-based face recognition model without sacrificing recognition accuracy. Our fundamental insight is that the computation increases exponentially with the depth of a network, the easily identifiable face images can be accurately recognized by the cheap features (pixel values at shallow layers), while the challenging samples that exhibit low quality, large pose variations or occlusions need to be processed by the expensive deep layers. The major contribution of this paper is the Reinforcement Learning Agent (RLA), which is proposed to learn a decision policy determined by a reward function. The policy adaptively decides whether the recognition should be performed at an early layer with a high recognition confidence, or proceeding to the subsequent layers, thus significantly reducing feed-forward cost for the easy faces. According to the extensive experiments on the popular face recognition benchmarks, Fast-FAR reduces the inference time by 14.22%, 20.61%, and 7.84% on the dataset LFW, AgeDB-30 and CFP-FP, respectively.