The Research On Face Recognition Based On Deep Convolutional Neural Network

The traditional face recognition model is limited by artificial feature extraction method,and the accuracy rate is low.In recent years,the deep structure of the convolutional neural network in the field of computer vision has made great progress,especially in the face recognition task.Convolutional neural network strong feature extraction ability,can overcome the human face light,attitude and so on.However,face recognition is a complex process,convolutional neural network also has some shortcomings.It is very important to study how to further improve the problem of convolutional neural network and apply it to face recognition task and solve practical problems,has very important significance.This paper is mainly based on convolutional neural network,discusses the shortcomings of deep convolutional neural network and its training strategy,and then study the deep convolutional neural network regularization method,finally combining the training strategy,the improved regularization method and measure learning method further improved facial feature extraction based on convolutional neural network model.The main work is as follows:1.A training strategy based on transfer learning is proposed to solve the problem of gradient vanishing in the training of convolutional neural networks for large numbers of samples.Knowledge of transfer includes the class distribution of samples and low-level features of source models.The class distribution provides the inter class correlation information of samples and indirectly increases the number of samples.Low-level features are general and can prevent gradient vanishing.Then we use the two parts of knowledge to pre-training the target model,and finally fine-tuning it with the real labeled samples.Experiments show that the training strategy accelerates the convergence speed of the model and improves generalization performance.2.In view of the existing the over-fitting phenomenon in convolutional neural network,proposed a kind of sparse Dropout regularization method,this method introduces sparse constraints on nodes in training,according to the value of the node activation to delete the network's nodes,in order to enhance the resistance capability of over-fitting.When testing,all the deleted nodes are restored,and the parameters of the training are retained to achieve the purpose of combining multiple local networks.Experiments show that the method of combining sparsity with Dropout has better generalization ability than traditional methods.3.In order to overcome the shortcomings of the Euclidean distance measurement method in facial feature expression,a face feature extraction model based on KL distance is proposed.The convolutional neural network is used to transform the input sample into a probability distribution,the probability distribution of measurement differences between different samples with KL distance,and define a cost function to optimize the parameters of the last distance,using a modified back propagation algorithm of convolutional neural network,the network has stronger ability to distinguish facial features.Experimental results show that this method can not only improve the accuracy,but also have better generalization performance.