Enhanced Representation Convolutional Neural Network Based Research On Face Anti-spoofing Methods

In recent years,with the rapid development of communication equipment and deep learning,face recognition technology has been widely used in places such as smart phones,security checks,and access control.However,the existing face recognition system exposes serious security risks when facing face image spoof attacks such as photo printing,video playback,and 3D masks.Aiming at the problem of face anti-spoofing,this paper conducts in-depth research on the convolutional neural network of enhanced representation.The main work is as follows:(1)A face anti-spoofing method(NSEAN)based on enhancement of non-significant information is proposed.This method uses the attention mechanism to enhance the model’s attention to feature non-significant areas,and retains more attacking features that are easy to be lost.The method also uses a multi-scale optimization strategy(MSRS)to obtain information of different granularities.Experimental results on multiple public data sets verify the effectiveness of the proposed method.(2)Propose a face anti-spoofing method(CERN)based on cascaded enhanced representation.The method includes two modules: cascaded enhanced feature extraction(CEFE)and cascaded enhanced feature fusion(CEFF).The cascaded enhanced feature extraction module combines the multi-level features through the backbone network,and at the same time uses the cascaded enhanced input space to encode the multi-level features again.Concatenated Enhanced Feature Fusion(CEFF)transfers high-level semantic information layer by layer to the lower layer,so that the low-level features contain both fine-grained attributes and high-level semantic information.In addition,this method also proposes a weight learning strategy to further enhance the discriminative power of predicting binary mask images.Experimental results on multiple public data sets verify the effectiveness of the proposed method.(3)Propose a face anti-spoofing method(Meta-CERN)based on meta-learning cascade enhanced representation.This method adds multi-label supervision of presentation attacks on the basis of CERN.By treating different types of presentation attacks as different labels,constraining the class spacing between presentation attacks in the source domain,and guiding the model to learn the characteristics of the same type of presentation attacks in different domains.In addition,a meta-learning method is introduced to divide the network into Metatrain and Meta-test to further enhance the cross-domain capability of the model.The test results on the face anti-spoofing cross-domain experiment verify the effectiveness of the proposed method.