Research On Face Recognition Algorithm Based On Deep Learning

This paper focuses on the face recognition algorithms based on deep learning,which has great academic value and application value.At the beginning,the paper analyzes the performances of different convolutional neural network(CNN)on face recognition.Then,to address the problems of some existing algorithms,three kinds of face recognition algorithms based on deep learning are proposed.The main contents are as follows.Firstly,a face recognition algorithm based on local binary pattern(LBP)and CNN is proposed to solve certain problem of face recognition being sensitive to illumination.LBP is a texture description method and it has good robustness to illumination.Therefore,we combine the grayscale image with the corresponding LBP feature map to generate a robust fusion image which is more informative than any of the input images.After that,the fusion image is fed to the CNN.Experiments show that the proposed algorithm can effectively reduce the influence of illumination on face recognition.Secondly,a face recognition algorithm based on the repetitive bilinear CNN model is proposed for addressing the problems that the bilinear CNN model has large amounts of parameters and high dimensional bilinear features.The newly proposed model uses InceptionResNetV1 instead of VGGNet to reduce model parameters.Simultaneously,bilinear operations are used twice to reduce the dimension of the re-bilinear features.Experimental results show that the repetitive bilinear CNN model can effectively improve the face recognition accuracy while reducing model parameters.Thirdly,an Angular-center loss function is proposed to solve the problem that the Asoftmax loss function has insufficient constraints on the separation of inter-class features.The proposed loss function combines the A-softmax loss and the improved center loss,so it has a double constraint effect on the features.After training,the feature is not only close to the weight vector,but also aggregates to the intra-class center feature.Experiments show that the Angularcenter loss function can obtain the deep features with inter-class separation and intra-class compactness as much as possible.Finally,it achieves 98.01% and 91.85% accuracy on LFW and YTF,respectively.