SC-NET: Spatial and Channel Attention Mechanism for Enhancement in Face Recognition

This paper proposes a spatial and channel attention mechanism module called SC-NET which is a lightweight yet effective method for deep convolutional neural networks. Recently, channel attention mechanism has been researched extensively and proved to be efficient in improvement of performance. However after carrying out rigorous empirical analysis, we find that channel attention and spatial channel attention improve the network's performance more efficiently. Therefore we incorporate both spatial information and cross-channel interaction in our SC-NET architecture. SC-NET is validated through extensive experiments on CASIA- WebFace and VGGFace2 datasets. By comparing our SC-NET with other methods, SC-NET has the best performance. Then when we apply our SC-NET to FaceNet(A Unified Embedding for Face Recognition and Clustering), FaceNet with SC-NET has achieved higher recognition accuracy than the original FaceNet and has reached state-of-the-art performance.