Research On Face Recognition Based On Convolutional Neural Network

Image processing is an important method of computer vision.In the research and application of images,face recognition technology has gradually matured.Convolutional neural network technology is also widely used in face recognition.The loss function module in neural networks is an important part of convolutional neural networks for face recognition.In order to improve the robustness of the recognition algorithm,this paper proposes a face recognition algorithm based on adaptive weighted fusion of multiple loss functions of convolutional neural network and a partial face recognition method of full convolutional network based on sparse non-local regularization weighted coding.The former recognition algorithm can improve the efficiency of complete face image recognition,and the latter can improve the recognition rate of partial face recognition.Because some face image is incomplete,the prior insufficient information greatly reduces the face recognition rate.Partial face recognition method of full convolutional network based on sparse non-local regularization weighted coding uses full convolutional neural network to obtain facial features.This method can receive pictures of any size for training and learning,and better save people Face information.According to the characteristics of local sparseness and non-local self-similarity,non-local regularization weighted coding is introduced into the sparse representation classifier.Compared with the existing algorithms,the proposed algorithm has higher recognition rate and better robustness on the public database.The work of this paper mainly includes the following three aspects:(1)Considering the important role of the loss function in the deep neural network used for recognition and the different emphasis of different loss functions,a weighted fusion method of multiple loss functions is proposed to solve the uncertainty and contingency of cognition by combining the idea of weighting.This method can effectively shorten the convergence time and improve the accuracy of recognition.(2)Based on sparse and non-local self-similarity,this paper introduce non-local regularized weighted coding in the sparse representation classifier.In this way,more detailed information can be retained,and errors between blocks can be better displayed.(3)By adding adjustment parameters,the dynamic matching process of images of different sizes is optimized to improve the errors generated in the dynamic matching process,so as to achieve the effect of making the classes close and away from each other.