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