Research On Face Recognition Based On Deep Learning Algorithm

Face recognition is a classic problem in the field of image.To solve the problems of low accuracy and rough feature point estimation,a face recognition method based on r-cnn algorithm is adopted.Characteristics of the method of choose and employ persons face detector effectively extract the facial features,at the same time,Res Nets-convolution neural network(R-CNN)combined with feature face algorithm is used for 2D face recognition.400 target faces are collected and mixed with 1000 sample faces in face database to form new face data set.The main work of this thesis is as follows:1.Aiming at the over fitting problem of CNN network,this dissertation proposes a network optimization method,improves the design of full connection layer,improves the original droopout random sparsity,and divides its output value into two levels(the former 1 / 2 and the posterior 1 / 2)from high to low in the training phase.The sparsity settings of the two levels are also different(the output value is low,the sparsity rate is high,and the output value is high,and the sparsity rate is low),so as to improve the whole connection.The improved CNN network is implemented in face data set,and the accuracy and loss of the network are given.2.Aiming at the shortcomings of r-cnn network,a strategy of replacing sliding window by fixed size segmentation is proposed.Because the size of sliding window is fixed,and it will cause window redundancy,but fixed size segmentation will produce target estimation with various sizes.Even if the face image is collected in extreme cases,the detection result is better.At the same time,the Convolutional Block Attention Module is adapted to the resnets network,which can enhance the feature extraction ability of the network and make the network more robust.The improved dropout technology is also added to the improved r-cnn network.3.A set of face detection and recognition software based on the improved r-cnn network is designed and implemented.The results show that it performs well in the unrestricted face data set,with high recognition rate and strong robustness.