Based On Temporal Shift And Depth-guided Attention Two-stream Network For Face Anti-spoofing

The current face recognition technology has been able to accurately identify whether the visitor exists in the base database,it is still difficult to distinguish the authenticity of the visitor.That is,whether it is directly accessed by the person or by someone else posing as the individual's information for access.Therefore,face anti-spoofing technology plays an important role in ensuring the security of the recognition system.Previous work is usually based on deep learning methods to extract semantic features in images,and estimate the depth information of human faces as spatial auxiliary supervision.Although the current methods based on deep learning can extract the semantic features in the image and use the face depth information and temporal information as auxiliary supervision,there are still four shortcomings.First,only the mean square error loss is used to constrain the training of the face depth prediction network,and the prediction results still need to be improved.Second,they focus on the deep semantic features,ignoring the shallow fine-grained features.Third,the attention mechanism is effective for solving the problem of face anti-spoofing,but the current method does not consider the characteristics of the face anti-spoofing task and ignores the scale information.Fourth,additional complexity will be introduced when modeling the temporal network,which is not conducive to the deployment of the model.In order to solve the first three problems,this paper proposes based on depth-guided attention two-stream network for face anti-spoofing DANet,which includes a two-stream network composed of depth estimate network DENet and multi-scale feature extraction network MSNet,and a depth-guided attention module DAM.The DENet adopts a codec structure and supervises the network training through the depth symmetric loss function proposed in this paper to more accurately estimate the depth information.The MSNet can extract the features of objects at different scales,and solves the problem that shallow features are ignored.The DAM uses the depth information of the face to guide the learning of the attention module and help the network focus more on useful scale targets.In order to solve the fourth problem,based on the temporal shift module TSM,this paper introduces TSM into the DANet,and further proposes the based on temporal shift and depth-guided attention two-stream network for face anti-spoofing,which can effectively extract temporal features without increasing additional calculation costs and model parameters.This paper conducted experiments on five public datasets: NUAA,CASIA-MFSD,Replay-Attack,OULU-NPU,and SiW.Through visualization and comparative analysis,the effectiveness of this algorithm is verified.