Research On Neural Network Algorithm In Face Recognition System

Stricter requirements are put forward for identity authentication in the information exchange process,with the development of computer technology and the popularity of the Internet.Face recognition belonging to Biometrics has been an important research topic in the field of identity authentication,because of its characteristics of non-contact recognition,stability,high efficiency,and low system hardware cost.Face recognition usually includes geometry method,feature face method,templates matching method.With the research and development of artificial neural networks,the face recognition method based on neural network has become one of the research hotspots in face recognition.The research of neural network algorithm applied in face recognition mainly starts from two aspects.Backward Propagation(BP)Neural Network.BP neural network is a common neural network used in face recognition.A face recognition system is designed in this paper,combining Principal Component Analysis(PCA),Histogram Equalization,Discrete Wavelet Transform and BP neural network.In order to study the performance of BP neural network in face recognition system,two BP neural network classifiers constructed.They are classifier-1 and classifier-2.It is found in this kind of face recognition system,through the simulation of face recognition systems with different BP neural network classifiers.The feature extraction method has a great influence on the face classification ability of BP neural network.And,the generalization ability of BP neural network.is poor,that means the accuracy of face recognition for untrained is much lower than that of trained faces.However,the classification ability and generalization ability of the BP neural network are improved,when the output of the BP neural network has been turned into discrete,that refers to sparse labels of training data.Convolutional Neural Network(CNN).First,a CNN based on LeNet-5 model has been trained face recognition in this paper,using the CMU_PIE face database.Then,the activation function and network structure of the CNN are optimized and improved for some problems including overfitting,slow convergence and low recognition accuracy,and a new LeNet-FC(LeNet Improved Face Recognition,LeNet-FC)convolutional neural network model is proposed in this paper.An activation function named Logarithmic Rectified Linear Unit(L-ReLU)is used in it,which is optimized and proposed in this paper.In terms of structure,the number of the new model's convolution layer is increased comparing with the CNN network based on the LeNet-5 model,the convolution kernel is narrowed,and the Dropout technology.is adopted.Finally,after making the target output of the CNN,that means the sparse training data labels,the LeNet-FC model's face recognition accuracy of test data reaches 99.85%in the face recognition training.Compared with CNN based on LeNet-5 model,the LeNet-FC model has better face recognition ability and strong generalization ability.Finally,face recognition system is designed in this paper,which is based on the LeNet-FC convolutional neural network model.The system performs facial feature extraction by the convolutional layer and pooled layer of the LeNet-FC,and realizes face recognition by comparing the Euclidean distances of features.According to the system's accuracy rate of face recognition reaches 96%in the simulation measurement using ORL face database,it is found that the LeNet-FC model has better performance in the actual face recognition system.Finally,the results of this system different performance simulation tests on ORL face database,MIT face database and AR face database show that The system has good recognition ability because of its strong robustness to differences in light illumination,micro-expressions,and minor gestures.