Attention-Based Spatial-Temporal Multi-Scale Network for Face Anti-Spoofing

Abstract

Face anti-spoofing is essential to the face recognition system. Although previous work has made great progress in spatial auxiliary supervision and temporal feature extraction, there are still great challenges in exploring the characteristics of facial depth as auxiliary information and modeling lightweight temporal networks. To address the above problems, we design a two-stream spatial-temporal network to explore the potential depth information and multi-scale information respectively. In the two-stream network, we introduce a temporal shift module, which can extract temporal information without additional calculations. In the depth information estimation network, we propose a Symmetry Loss for more accurate auxiliary supervision. Then, we construct a scale-level attention module based on the estimated depth information which fusions the information from the two-stream network to extract the essential discriminative features. Finally, a fully connected network is followed to judge the authenticity of the image. Experiments show that our method can get impressive results on five datasets including NUAA, CASIA-MFSD, Replay-Attack, OULU-NPU and SiW.