Research On Convolution Neural Network Modeling Methods And Application In Face Recognition

In recent years,convolutional neural network(CNN)has made a breakthrough in the field of computer vision,and with its special network structure and strong ability to extract features,the accuracy of face recognition can be improved by leaps and bounds.The selection of face database and the design of deep face recognition network are the key factors to improve the face recognition rate,because the face recognition research is greatly affected by the factors such as face pose illumination occlusion expression and so on.This paper mainly studies the application of CNN model in the field of face recognition in the natural environment,and makes a deep research on expanding the face database,constructing and training a reasonable depth face recognition network model,and builds a classroom attendance system.The research contents of this paper are as follows:(1)In view of the disadvantages of sigmoid,tanh activation function and ReLU forced sparse convolution neural network,the nonlinear,differentiability and monotonic properties of activation function are combined.The activation function is improved by combining the property of ReLU sparsity and the smoothness of sigmoid function.The improved activation function has the advantages of simple derivation of ReLU function and solution of the disadvantage that ReLU function makes neural network too sparse and leads to serious lack of information.(2)The network model of depth face recognition is constructed by the idea of residual learning.In view of the complex network structure of DeepID face recognition and the long training time of LightCNN,the network model of face recognition designed in this paper can reasonably reduce the network parameters.The training time of convolution neural network is reduced by using xavier and msra initialization method which is better than random initialization method,and 99.22%accuracy is obtained on LFW(Labeled Faces in the Wild)face database after pretreatment,and the twin network(Siamese Network)is also studied in depth.(3)Aiming at the shortcomings of the existing loss function training,the convolutional neural network model is not compact in class and separated from each other when extracting face image features,a loss function with better performance is proposed.By introducing feature normalization,weight normalization and increasing punishment for correct classes,the convolutional neural network model obtained by training has the characteristics of compactness within classes and separation between classes when extracting features.(4)Finally,the trained depth face recognition network is applied to the actual scene to realize the pattern classroom attendance system.The camera is used to collect photos in the classroom,and the multi-task cascade face detection algorithm(MTCNN)based on convolution neural network is used to get a single face image,so as to construct the classroom attendance face database as the training data of convolution neural network.It improves the ability of the convolution neural network to extract the features of the face images in the complicated environment in the classroom.Attendance system extracts features from three-channel color images,thus preserving the complete information of face images.The system verifies the effectiveness and practicability of the method in this paper,and meets the requirements of the application of attendance assessment in class.