Research Of Face Recognition Algorithm Based On Improving Manifold Learning And Deep Neural Network

As an indispensable identification technology in the field of modern information security,Face Recognition(FR)has become one of the most popular target recognition technologies,which relies on facial differences,and has been widely used in the fields of surveillance and security,person-to-person comparison and so on.With the development of Deep Learning,the performance of face recognition has achieved a qualitative leap.However,due to the influence of sample size,illumination and other conditions,the Deep Neural Networks algorithm can not deal with the small sample size problem very well.Based on the fact that human face conforms to high-dimensional manifold,Manifold Learning algorithm can better preserve the useful information of human face in the process of dimension reduction.Considering the advantages of the two algorithms,an algorithmic thought based on improved manifold learning and deep neural network is proposed.The main work is as follows:Firstly,the two-dimensional Locality Preserving Projection(2DLPP)algorithm is studied.Although the linear dimensionality reduction algorithm can preserve the local manifold structure of face,which is a high-dimensional feature,it lacks the consideration of category information.To overcome this shortcoming,an improved face recognition algorithm of two-dimensional locality preserving projection is proposed.This method introduces class weights to minimize intra-class differences and maximize inter-class differences simultaneously,and uses the idea of two-directional linear dimension reduction feature extraction algorithm for reference.It combines two-dimensional principal component analysis and two-dimensional linear discriminant analysis algorithm in row direction and column direction respectively to extract useful information from face images to a greater extent.This can solve the dimension disaster and the small sample problem very well.Secondly,we analyze the good performance of existing Gabor features to face images,and its multi-channel features can improve the recognition rate,but the multi-channel face representation will cause a large amount of computation when classifying.To overcome the shortcomings of multi-channel Gabor Face representation algorithm,we propose a face recognition algorithm that combines MGFR and improves two-directional face feature dimensionality reduction.By selected from scale and direction,the new Gabor face representation is constructed by the optimal channel combination method,and is used as the input of the improved two-dimensional feature reduction algorithm,so as to maximize the useful feature information on small sample data sets to improve the effectiveness of the algorithm.Finally,the BP neural network and RBF neural network for classification are introduced,and the PCANet and 2DPCANet models with linear dimensionality reduction are introduced.Aiming at the current application of neural network in image classification and feature extraction,the algorithm combining feature dimensionality reduction with neural network is discussed.The features extracted from dimensionality reduction process are used as input of neural network,and classical BP neural network and RBF neural network are used to classify,so as to obtain higher classification accuracy.