Research Of Face Recognition Method Based On Convolutional Neural Network

Artificial intelligence is a current hotspot in internet giants and research institutes.Face recognition is a kind of artificial intelligence which widely used in various fields.In the era of big data,the advantage of artificial neural network on face recognition is becoming more and more obvious.Convolutional neural networks developed on the basis of artificial neural network are a kind of supervised learning network,which combine the depth learning structure and use back propagation algorithm with gradient descent to train network.It integrate features extracttion automatically and classiffier to reduce the complexity of the face recognition process.The convolution layer with sparse connection and weight sharing can reduce the network parameters greatly that need to be trained and improve the network training performance.The subsampling layer can preserve the feature data while it can reduce the image resolution to optimize the network structure.In this paper,the iterations number of network training convergence and the recognition rate are studied by using three different convolutional neural network from simple to complex,which change the number of network layers and neurons based on the LeNet5 structure.The experimental results show that the more the complex network,the network performance not necessarily the better,the error of the test set is higher than that of the training set,which indicates the stability of the convolution neural network needs to be improved,and the best network structure of different databases is also different.Too large or too small learning rate of training network can make the convolution neural network not to converge,so three kinds of different learning rates with different changing trends are proposed to train three different network structures on ORL face database.The experimental results show the trend curve that the beginning of the change is slow and then slowly increase is more suitable for training.The traditional convolution neural network is generally selected by the artificial experience of the best performance structure,not only resulting in heavy workload but the generalization of structure is poor.Aiming at the poor generaliaztion and universal property,an adaptive convolution neural network algorithm is proposed to determine the network structure automatically without performance comparison,and it is compared with the traditional convolution neural network on three different face databases.Experiments show that the adaptive convolution neural network achieves good balance between training time of network and test error,and also improves the stability of the network.In addition,this paper comprehensively analyzes the influence of the branch network structure,the network expansion interval and the network expansion factor parameters on the network performance of the adaptive convolution neural network algorithm on ORL face database.