The Research Of Face Recognition Based On Deep Learning

As an important part of artificial intelligence,target recognition has made amazing development in recent years.As a branch of target recognition,face recognition is more and more widely used.However,face recognition is easily affected by internal and external differences.Face recognition methods are difficult to achieve the desired results.As a new field in machine learning research,deep learning is different from traditional shallow network.It has better expressiveness and greater generalization ability.It can simulate the working mechanism of the human brain,building deeper network models and training on large-scale data,realizing the approximation of complex functions,obtaining robust facial features,and effective improve the recognition effect of images,so that the accuracy of face recognition is greatly improved.The main work is as follows:(1)Introduce the basic theory of deep learning.The gradient descent algorithm,forward propagation algorithm and backpropagation algorithm required for training deep neural networks are introduced.Next,the convolutional layer,the pooling layer and the fully connected layer of the convolutional neural network,which is one of the representative algorithms of deep learning,are explained.Finally,the overall structure of the network is introduced through a typical convolutional neural network,LeNet-5.(2)A nine-layer convolutional neural network was designed.Use the dropout waiver technology,introduce L2 regularization penalty items,and use sample data expansion to solve network over-fitting problems.The network feature extraction is accelerated by transforming the fixed learning rate into variable learning rate,the combination of batch gradient descent and random batch descent method,new momentum acceleration method and cross entropy loss function instead of quadratic loss function.Finally,the support vector machine(SVM)is selected as the classifier to classify the extracted features.(3)Part of the data in the Casia-Webfaces dataset was selected to pre-train the designed nine-layer convolutional neural network,and the FERET dataset was selected as the target dataset to formally train the network.In the end,a better recognition rate was obtained in a shorter period of time.In comparison with other face recognition methods,the method of this paper is superior in training time and accuracy.In the text,there are 34 pictures,9 tables,and 54 references.