Research On Face Recognition Algorithm Based On Convolutional Neural Network

In recent years,face recognition has made great progress,which is attributed to the rapid development of deep neural network,especially the development of convolutional neural network.Convolutional neural networks have good robustness in face recognition,which is suitable for multi-scene recognition and can effectively reduce the recognition rate decrease due to light and angle to some extent.Convolution neural network plays an important role in face recognition.This thesis takes face recognition algorithm based on convolution neural network as the research topic,focusing on deep network,loss function and other problems in face recognition.A network of effective feature extraction is of great importance for face recognition task.This thesis studies and analyzes the classical convolutional neural network.Based on the classical residual neural network,the squeezed excitation module is integrated with the residual network.Compared with other convolutional neural networks,the improved residual network can improve the accuracy of face recognition.By comparing the effects of different feature extraction layer Settings on accuracy,a reasonable feature extraction layer was selected for face recognition network.This thesis introduces a variety of classical loss functions for face recognition and analyzes their advantages and disadvantages.In view of the fact that the similarity between the eigenvectors learned by the existing loss functions when the training data are not uniformly distributed cannot accurately represent the similarity between their underlying faces,the original angular distance loss function is improved and other attributes are used to regulate the learned feature mapping.The improved loss function can learn more discriminating face features and improve the accuracy of face recognition.Experimental results show that using the CASIA Webface training dataset,with the modified residual network for face feature extraction,based on attribute driven loss function to guide the network training,through a large number of analysis and experiment,this algorithm in LFW data sets on face recognition accuracy reached99.67%,in face recognition tasks and face Mega Face verification tasks for face recognition accuracy of 74.531% and 87.134%,which shows that the algorithm has a good effectiveness.In order to verify the effectiveness of the algorithm,a face recognition system based on the face recognition algorithm in this thesis was finally implemented,which integrated other key modules such as face detection,feature point detection and face feature comparison in face recognition tasks.The test in the actual environment shows that the algorithm has good robustness to attitude and illumination,as well as high recognition accuracy,which not only proves the feasibility of the algorithm in theory,but also verifies that the algorithm in this paper performs very well in daily application scenarios.