Research Of Face Recognition Method Based On Deep Neural Network

With the growth of artificial intelligence and deep learning,face recognition is gradually expanding in many industries and fields.Face recognition technology based on convolutional neural network is widely used in security,business,finance and entertainment fields.Compared with the traditional face recognition method,the model can not only automatically extract image features,but also automatically learn to get a higher recognition rate.According to the size of training data set,this method designs a suitable network structure.This paper improves and optimizes AlexNet network structure to improve training efficiency and stability.The main research contents are as follows:(1)Convolutional neural network is a deep and sparse traditional artificial neural network.The model structure,forward and back propagation algorithms of traditional artificial neural network are analyzed firstly.Then the convolutional layer,excitation layer,pooling layer and full connection layer are described.Then the forward propagation and back propagation are deduced by mathematical formula.Then the training process is described.Finally,the gradient descent algorithm is briefly introduced.(2)Face image preprocessing includes face image detection,face image clipping,face image standardization,classification marking,data conversion,training set average file,etc.,and face recognition algorithms based on the machine vision library OpenCV feature face,linear discriminant analysis,local binary mode histogram are implemented.(3)To solve the problems of slow loss gradient descent,slow rate gradient rise and poor stability of AlexNet network in face recognition,a scheme based on improved and optimized network structure model is proposed.Firstly,the AlexNet network structure and related parameters of the model are analyzed,and the AlexNet network structure and parameters are improved.According to the VGGNet network model,the structure model of the improved convolutional kernel network is proposed.Then the network was analyzed and found that the error rate of LRN layer was 0.1%higher than that of LRN layer without LRN layer,so the LRN layer was deleted.Then,by testing the influence of the number of fully connected layers of AlexNet on face recognition,three groups of deep neural network face recognition experiments were carried out.The number of fully connected layers was set as 1,2 and 3,respectively.The results showed that the accuracy of all three training sets was 96.25%,that is,the accuracy of deep learning on the training set had no impact.However,when the full connection layer is 1,the recognition rate of test set is 96.25%,and the other two are97.5%.When the full connection layer is 2,the recognition rate of the sample set is99%,and the other two are both 98.75%.According to the data obtained from the test,the performance of face recognition is the best when the number of full connection layers is 2.The improvement plan is to replace the original 11×11 single large convolution kernel with the superposition form of 7×7 and 5×5 small convolution kernel in the first convolution layer,delete a full connection layer and LRN layer,and debug the structure with existing parameters.Finally,according to the GoogleNet network model,the optimization scheme is proposed,which is partially replaced by two small convolution kernels of 3×3 and 1×1 in the second convolutional layer 5×5convolution kernel,and is denoted as the optimized improved convolutional kernel network,and then the parameters are adjusted.The improved and optimized network structure is tested on face data sets such as ORL,GT and Faces95,respectively.Comparing the performance of the three networks,the results show that AlexNet has slow learning speed,low recognition rate,large loss value and poor stability.The improved convolutional kernel network improves the learning speed,face recognition rate,stability and loss of face data.However,at some iteration points,the change is stronger than AlexNet.The late loss values of GT and Faces95 face data set are both slightly higher than AlexNet,and the recognition rate of Faces95 is also slightly lower than AlexNet.The optimized and improved convolutional kernel network solves these problems.