Research On Face Anti-spoofing Algorithm Based On Dual-stream Network

In facial recognition applications,criminals use facial photos or facial videos to impersonate a living human face to commit crimes.In order to solve this safety hazard,this paper studies the face anti-spoofing,aiming to distinguish the live face and the nonliving face collected by the camera.If the classic image binary classification algorithm is used,there are fewer distinguishable features between live faces and non-living faces,the model is easy to overfit,and it is greatly affected by changes in illumination.Therefore,our method uses pixel-level supervision to learn deep features,and at the same time uses a lightweight network to extract texture features to build a dual-stream network to improve the accuracy and generalization ability of the face anti-spoofing algorithm.Aiming at the problems of the traditional image binary classification algorithm,which is easy to overfit in the face living detection task,and the illumination change has a greater impact on the model effect,this paper designs the G-Attention-Depth Net depth estimation network.First build a benchmark depth estimation network;then propose a central gradient convolution,let the network use the gradient between pixels to learn the depth features,while extracting the edge features of the illumination,reducing the impact of illumination changes;finally,this paper designs an adaptive multi-scale feature fusion network to make full use of low High-end features.Compared with the benchmark network model,the average error rate of G-Attention-Depth Net on the NUAA dataset is reduced by 1.7%,on the OULU-NPU protocol 1,it is reduced by 3%,and on the OULUNPU protocol 2,it is reduced by 3.7%.The experimental results show that the improved depth estimation network in our method can effectively learn more discriminative depth features,and has strong detection capabilities for photo and video attacks.The depth label used by G-Attention-Depth Net is generated by the PRNet network.This generative depth label contains slightly less information than the real depth map,so the depth features learned by the model based on the generative deep label are not accurate enough.Based on this problem,this paper proposes the SC-MobileNetV3 texture feature extraction network.Taking Mobile Net V3 as the benchmark model,adding Sobel convolution to the bottleneck layer makes the network pay more attention to local texture details;at the same time,a central pooling layer is designed to replace global average pooling,making full use of the central area of the face image that has more than the edge area Information and higher importance characteristics.Finally,the two models of G-Attention-Depth Net and SC-Mobile Net V3 are merged to design a dual-stream network structure to make full use of complementary information.Compared with G-AttentionDepth Net,the average error rate of dual-stream network on OULU-NPU protocol 1 and 2is reduced by 0.2%.The experimental results show that the dual-stream network proposed in our method can effectively compensate for the defects of virtual deep tags and further improve the accuracy of face anti-spoofing.