Research On Face Liveness Detection Methods Across Scenes

With the rapid development of face recognition technology,its security is getting more and more attention.Face recognition system is mainly based on image information and vulnerable to photo attacks or video attacks.Face liveness detection technology to determine whether the face image is a living body,which has been an important part of the face recognition system.Many face liveness methods have been proposed,but they don't generalize well when across scenes.This thesis summarizes the main research results in the field of face liveness detection.Aiming at the generalize problem in face liveness detection,this thesis studies domain adaptation,public features and face depth information,and designs two kinds of face liveness detection algorithms.The main work of the thesis is as follows:1.A face liveness detection algorithm based on multi-layer domain adaptation is proposed,which mainly solves the problem of inconsistent data distribution between datasets.The texture information can be of great help in face liveness detection,because the difference information between the genuine face and fake face is small.At the same time,as the convolutional neural network deepens,the features extracted by the network will lose the texture information.Therefore,this thesis fuses the low-layer and high-layer features of the network.To improve the generalization of the model,this thesis applies the Maximum Mean Discrepancy on the fully connected layer in the network,which eliminates the distribution distance between the source and target domains.The half total error rate of this method on the two public datasets of Replay and CASIA is 31.2% and 33.5% in cross-scene,respectively.The experimental results show that our method further improves the generalization of the face liveness detection algorithm.2.The traditional face liveness detection mainly focuses on the difference information between the genuine face and fake face,which ignores the similarity between the datasets.Although the data distribution varies greatly between different datasets,there is largely public information between the same attack types in different datasets.Therefore,this thesis extracts the public features of the source domain and the target domain to train the classifier,so that the model has better generalization ability.Besides,the genuine faces have a three-dimensional spatial structure,while fake faces are often flat structure.To further improve the generalization of the model,this thesis uses three-dimensional face reconstruction technology.Face depth information is used as auxiliary information for the network.The experimental results show that public features and face depth information can greatly improve the generalization of the model.