Research On Face Recognition Method Based On Convolutional Neural Network

The rapid development of computer science and technology and computer hardware equipment makes face recognition technology based on deep learning widely used in people's daily life,such as video surveillance,mobile payment,smart attendance and access control systems.From the past to the present,various related algorithms have been proposed to solve various problems encountered in face recognition and made significant progress.In some cases,the face recognition rate of computers even exceeds that of humans.However,under unconstrained conditions,face recognition technology still faces many challenges,such as changes in face pose and lack of face image information caused by factors such as ambient lighting,low resolution,noise,and partial occlusion.These factors have greatly affected the effect of the face recognition process,and it is worthy of our further study.In addition,how to effectively reduce the large amount of computational cost brought about by neural network convolution operations is also worth exploring.In response to the above problems,this paper further studies the face recognition method based on deep learning.The main contents of the work are as follows:(1)Train the P network,R network and O network of MTCNN to realize face detection;use the DBSCAN algorithm to cluster and clean the original data set,and at the same time expand the face data set,remove the face background interference,and use the human face.The average pixel padding of the face.(2)Improve the residual network.Use Softmax loss and Center loss as the joint loss function to increase the inner distance of the face class,reduce the class separation,and optimize the activation function at the same time;introduce the SEnet channel attention mechanism to enhance the ability to acquire important features,and add depth residuals The difference identity mapping module realizes the function of mapping side face features to positive face features in the depth feature space layer,and further improves the face recognition rate.Finally,the effect of the algorithm is verified on multiple data sets of LFW,CFP,IJB-A,YTF and SLLFW.(3)Realize the lightweight design of the network,and replace the convolutional layer in the network with the Ghost module.First,standard convolution is used to generate a small amount of eigenfeature maps,and then the Ghost module is used to generate "ghost" feature maps from the eigenfeature maps,and finally they are spliced and output.The experimental results show that the use of the Ghost module can effectively reduce the amount of parameters and reduce the amount of calculation,while still ensuring a good face recognition effect on the network.(4)Complete the design of the dynamic face recognition system to realize the detection,recognition and tracking of video faces.