Research On Face Pose Reconstruction Based On Neural Network

In the field of face recognition,the pose change of face image has often led to the decrease of the recognition rate,which is also a bottleneck in the field of face recognition technology.Therefore,this paper takes the face pose reconstruction as the research goal,researches and improves the face pose reconstruction technology,and compares it with the existing method to judge its performance.By comparing and analyzing the different face pose reconstruction methods,the locally linear regression(LLR)algorithm has been chosen to be the basis of the developed method.Based on the linear regression of the deep neural network,the reconstruction method has been put forward.On the basis of the linear regression relation of LLR,combining with the constant factor in the deep neural network and adding a quadratic factor to preserve the nonlinear information of the image,the function of the face gesture reconstruction has been obtained.The input of the side face image image has been divided into many small parts,and then the collection of the small image has been put into the trained neural network to produce a colletion of reconstructed small piece of the front face.Then the corresponding front face can be obtained.In the process of implementation,after establishing the database of the face images,the database are preprocessed to facilitate the work of face pose reconstruction.Through the design of the neural network,the number and structure of the neural network have been determined,the weights of the convolution layer have been initialized,the appropriate incentive function has been selected,the learning rate of the parameter updating has been carried out.Finally,the design of neural network used in face pose reconstruction method has been completed.This paper has added the database to the neural network for the realization the face pose reconstruction,then has processed the face recognition of the reconstructed image.The results show that the method in this paper can efficiently improve the recoginition rate.