Research On Face Recognition Based On Deep Learning

Face,as an important feature information of a person,is unique and natural,so face recognition is widely used in the field of identity authentication.Deep learning has the characteristics of feature extraction and feature recognition similar to human brain mechanism.With the rapid development of computer performance in recent years,deep learning,especially large-scale deep learning neural network,has been realized.In deep learning,Convolutional Neural Networks(CNN)have a high degree of feature invariability for spatial transformations such as translation and rotation of objects,so they have unique advantages in the field of face recognition.However,face recognition under unconstrained conditions is often affected by various comprehensive factors such as attitude change,expression change,different degree of occlusion and exposure.Meantime,almost all of the deep convolutional neural networks have problems such as too many parameters and gradient diffusion during training.Therefore,based on the deep learning theory,this paper proposes a multi-loss function feature fusion algorithm model to improve the face recognition rate.This paper introduces and analyzes some representative face recognition algorithms at home and abroad,because the face features extracted by the convolutional neural network are more discriminative and convenient without the need to manually adjust parameters.Therefore,this paper extracts face features based on the convolutional neural network.However,since the convolutional neural network extracts the global features of faces,its features highlight the degree of difference between different faces rather than the changes in exposure,expression and other directions.Accordingly in order to improve the recognition accuracy,in adding the high-dimensional feature extraction local binary pattern(highdimensional local binary pattern,HLBP)to obtain the local characteristics of human face,local characteristics with different exposure,expression change not only has a certain robustness at the same time effective depicting the details of face facial expression change,to some extent compensate for the convolution of the insufficiency of the neural network in detail feature extracting change.In general,the deeper the layers of the neural network,the more discriminant the extracted face features are,but as the number of network layers increases,its training parameters will increase exponentially,so if we do not solve the problem of the number of parameters,the network will not only not achieve the final expected recognition effect,it is likely to lead to a significant reduction in recognition rate.because the gradient of the error is continuously reduced during the back-propagation of the error.If you do not take a suitable method to solve the problem,then the network The earlier the corresponding parameters are,the less fully trained they are,then the features they extract will not be discriminative.Therefore,the experiment in this paper refers to the structure of densely connected convolutional neural network,which can effectively solve the problems of excessive parameters caused by too many layers and gradient diffusion during training.This article will successively in CASIA datasets Web Face,LFW non-binding data sets and Mega Face experimental comparison and analysis on large data sets,the results show that both in the constraint conditions and the constraint conditions,in the same order of magnitude network under the recognition efficiency,best performance especially in the restricted conditions,good the superiority of the model in this paper,the proposed algorithm is verified...