Research On Silent Face Anti-spoofing Algorithm For Community Correction Objects

In order to improve the effectiveness of supervision,judicial authorities require community correction personnel to use mobile phones for daily check-in and ideological reporting on many occasions.However,problems of “man-machine separation” and “fake check-in” that are common in actual work.It is urgent to introduce face recognition into daily supervision at a limited cost under the condition of existing equipment.But a simple face recognition system for mobile phone users is extremely vulnerable to non-living attacks such as face printing photos,electronic photos or videos,and it needs to be combined with suitable face anti-spoofing(FAS)technology to greatly improve the security of face recognition system.Compared with other FAS implementation methods,silent FAS has been research hotspot due to the convenience of recognition process and the need of no additional dedicated auxiliary hardware.Since most existing algorithms regard FAS as a binary classification task,ignoring the difference between living face and spoofing face and the adverse impact of class imbalance.In this regard,this thesis proposes a supervised network model leaning scheme based on multi classification cross entropy loss to overcome the problem in binary classification and improve the accuracy of model for spoofing recognition;besides,a two-stream feature fusion network is constructed to integrate the gradient and color features of face images into the recognition model,so as to further improve the feature representation ability of the model.The specific research contents include:(1)The non-living face samples are subdivided into print attack and display attack,and the multi classification cross entropy loss supervision model is used for training.On the one hand,it can reduce the impact of data collection environment characteristics,and make the model pay more attention to the differences between classes and the common features within classes.On the other hand,it can balance the differences in the number of samples of each class,and further improve the training effect of the model.Experiments show that the multi classification strategy can improve the performance of the model,reduce the classification error rate and improve the generation performance.(2)In view of the difference of gradient and color features of face images caused by secondary acquisition,a two-stream feature fusion neural network is constructed to further optimize the recognition model based on face gradient and color features.Firstly,the image gradient operator is integrated into network structure by using depthwise sperate convolution,and the gradient neural convolution network is constructed for image feature extraction;Then,combined with the idea of attention mechanism,two different feature fusion strategies are adopted to adaptively fuse the feature vectors extracted from RGB and HSV/YCr Cb color space,so as to further improve the feature representation ability of the model.Experiments show that the two-stream network model can effectively reduce the classification error rate and improve the generalization ability.