Research On Face Recognition Algorithms Based On Multi-layer Non-negative Matrix Factorization Architecture

In recent years,artificial intelligence has become the focus of our daily life.In the era of artificial intelligence,face recognition technology plays an important role in our daily life,and had been widely used in the face unlocking and face payment of smart phones,so it is particularly important to improve the recognition rate of face recognition technology.For this reason,some advanced face recognition algorithms have been proposed and widely used,such as Non-negative Matrix Factorization(NMF)algorithm.NMF algorithm represents the original image as a linear combination of a set of basis images.This image representation method is in line with the idea of “parts constitute a whole” in human thinking.Due to the change of shooting environment,equipment,shooting angle and face occlusion,the facial images will always be non-linear distributed.However,as a linear algorithm,NMF has great difficulties in dealing with more complexly non-linear distribution problems,the basic images and features learned from the shallow NMF algorithm are relatively simple,which cannot reflect the deeper features of the data,while the human visual system adopts a hierarchical and nonlinear method when analyzing images.The objective of this dissertation is to remedy the aforementioned limitations of the above algorithm.This dissertation includes 5 chapters.The first chapter introduces some related research work of NMF algorithm.The contributions of this dissertation are concentrated in chapter 2 to chapter4.The following briefly introduce the main research work of these three chapters.,And the chapter 5 is the summary and prospect.As a linear algorithm,the classical NMF algorithm is not effective in dealing with non-linear distribution data.RBF neural network is an effective nonlinear learning model with strong nonlinear fitting ability.Hidden neurons and weights play an important role in neural networks.In Chapter 2,a neural network algorithm named Nonnegative Matrix Factorization based on Radial Basis Function(NMRBF)is proposed.NMRBF algorithm uses the main idea of NMF to train the parameters of RBF neural network.This algorithm can improve the accuracy of hidden neurons,and the iterative formula of weight can ensure its solvability and interpretability.Experiments will be carried out on ORL and Yale face databases in chapter 2.The experimental results show that the NMRBF algorithm proposed in this paper performances well.However,NMRBF algorithm is still a shallow algorithm,it is difficult to extract the deeper features of the data.To extract features reflecting the deep localization characteristics of images,a novel deep nonnegative basis matrix factorization architecture based on underlying basis images learning(UBIL)is proposed in chapter 3.This architecture learns the underlying basis images by deep factorization based on the basis images matrix.And this architecture has a strong interpretability and practical physical significance.To implement this architecture,chapter 3 proposes a Deep Non-negative Basis Matrix Factorization(DNBMF)algorithm and Regularized Deep Nonnegative Basis Matrix Factorization(RDNBMF)algorithm.The convergence of the algorithms is proved theoretically in Chapter 3.Experiments will be carried out on FERET,ORL,AR,CMU,Stirling and Yale face databases in chapter 3.The experimental results show that the deep non-negative basis matrix factorization architecture based on UBIL has better recognition performance than other deep factorization architectures based on coefficient matrix.However,both DNBMF algorithm and RDNBMF algorithm are linear algorithms,the recognition performance still needs to be improved when dealing with the data with more complex nonlinear distribution.The Regularized Deep Nonlinear Non-negative Basis Matrix Factorization(RDNNBMF)algorithm is proposed to handle pattern recognition tasks with more complex data in chapter 4.RDNNBMF algorithm projects the original samples into a high dimensional space by a nonlinear map,and deep factorization based on UBIL is performed on the mapped samples in this high dimensional space.Five face databases,namely FERET,ORL,AR,Stirling and Yale databases,are selected for evaluations.The experimental results show that the nonlinear algorithm RDNNBMF performs better than the linear algorithms RDNBMF and DNBMF,and so are the other nonlinear algorithms.