Research Of Face Recognition Based On Deep Convolution Neural Network

Big data contains great development in recent years.Combined with the continuous improvement of high-performance computer,it is to bring the rapid development of machine vision technology.Deep convolution neural network is also developing quickly,and its main applications include speech recognition,image processing and Natural Language Processing,etc.In this paper,the application of convolution neural network model in face recognition in static environment is studied in the current environment.Compared with the traditional face recognition method,the convolution neural network model does not need to manually carry out a large number of complex feature extraction algorithm design,only need to design a feasible network model,and load a large number of face training data set to the network model,and then use automatic training,so that we can get a good recognition rate.The model is saved,then it is an end-to-end face feature extractor.Although the method is simple to operate,it is necessary to design a reasonable network structure according to the training data set,and the most difficult key point is the adjustment of the hyperparameters and the design of the optimization algorithm.Therefore,this paper constructs two network models which match the computing resources and data resources with the residual network and the fusion network.By repeating the adjustment of the hyperparameters and the debugging optimizer,it can converge on the training set and finally get a better identification rate.The main contents and innovations of this paper are as follows:1 introduces the basic theoretical knowledge of convolution neural networks.Convolution neural network is essentially a deep and sparse traditional artificial neural network.Therefore,this paper begins with a detailed analysis of the model structure,forward and backward propagation algorithms of traditional artificial neural networks.And then the content transitions to convolution neural networks,and its important components,such as convolution layer,excitation layer,pool layer and all connection layer,are summarized.Finally,the note of the convolution neural network training is described.2 The system architecture and programming model of the TensorFlow system are described,and the face data is preprocessed,including face detection,data enhancement,image standardization and face center loss.3 The face recognitions based on improved MyVGGNet and MySqueezeNet networks are proposed.Firstly,the network structure and the related parameters of model VGGNet-16 and SqueezeNet are analyzed.Then,The paper proposes that the network structure and parameters of the original VGGNet-16 and SqueezeNet are optimized.Batch normalization layer is added between each convolution layer and the activation layer.At the end of the VGGNet-16 network,one 1 x 1 convolution layers is used instead of the three fully connected layers,and the global average pool layer is added,a new MyVGGNet and MySqueezeNet models are gotten,the final two models not only solve the original model of high hardware requirements,limited data volume has been fitted and other issues,but also successfully applied to the task of face recognition.The original networks are transferred,so that it is easier to obtain the optimal solution in the face recognition task.Finally,in the LFW data set the correct rates of 94.3%and 95.1%is obtained respectively.4 Two network models BTreeFuseNet_v1 and BTreeFuseNet_v2 with depths of 22 and 19 are constructed by combining the residual network and integrating into the branch parallel,fusion and cascade three structures and adopting ReLU function,BN layer,Dropout layer,xavier method and truncated Gaussian function initialization method,Adam optimizer and other skills,and the basic structure,the whole architecture and the model parameters of the network model are described in detail.Finally,through continuous training in the FaceScrub data Set and adjusting the network hyperparameters,a better model is obtained,and then 94.9%and 95.5%accuracy are achieved in the LFW test set for face verification,respectively.