As an indispensable part of today’s social development,image recognition not only facilitates people’s daily life,but also provides new thinking for the development of science and technology.In the application of image recognition,the image data is often very large,and will face the problem of "dimension disaster"。Thus,an approach of data compression to reduce the dimension is essential to solve the problem.Non-negative Matrix Factorization(NMF)can compress and reduce the dimension of the data,and can extract the features of the image data.Graph Regularization Non-negative Matrix Factorization(GNMF)has the ability to add graph theory and manifold hypothesis,as well as consider the geometric structure of data.However,there are still some problems in GNMF,such as data redundancy,unable to accurately extract image features and poor adaptability of the algorithm in different image sets,which lead to low image recognition rate.Therefore,it has significant research value to carry out the research of image recognition based on Graph Regularization Non-negative Matrix Factorization.In this dissertation,manifold learning and Non-negative Matrix Factorization are applied to image recognition:1.Discuss the basic process and characteristics of image recognition in theory,and master the basic technology of image recognition.This dissertation explores the application rules of GNMF method,research on how to improve it and summarizes and classifies the improved GNMF methods for image recognition,so as to lay a foundation for designing efficient image recognition methods.2.Improved Graph Regularization Non-negative Matrix Factorization(New-GNMF),filter redundant information,extract image features accurately,and expand the application scope of image recognition.Firstly,NMF is combined with the graph theory and manifold hypothesis to extract the main features of the image data locally and preserve the intrinsic geometric structure of the data.Moreover,the threshold S is set for the decomposed base matrix,and the threshold judgment is used to optimize the processing,filter the redundant data,reduce the interference,and effectively solve the problems of GNMF in the actual image recognition process,such as a large number of redundant data,unable to accurately extract image features and so on.Then,the objective function and iterative formula derivation process of the improved GNMF are given.Finally,New-GNMF is applied to image recognition,and image recognition simulation experiments are carried out on the four image databases of COIL20,PIE-pose05,PIE-pose09 and YaleB.The Accuracy(AC)and Normalized Mutual Information(NMI)of clustering recognition are calculated respectively,The performance of the improved algorithm is evaluated using the evaluation system.The results show that the approach can significantly improve the image recognition effect.3.Improved Sparse Graph Regularization Non-negative Matrix Factorization(New-SGNMF),threshold optimization and norm constraint matrix are set to further improve the effect of image recognition.Because that the sparsity of image recognition based on New-GNMF can’t be controlled in the process of recognition,New-SGNMF makes full use of the inherent geometric structure of image data,optimizes the base matrix in two steps,filters the redundant information,and controls the sparsity at the same time.Firstly,a threshold value S is set to judge the threshold value of the decomposed base matrix to filter the redundant information in the data.Then,L2 norm is used to implement sparse constraint on the basis matrix,which is integrated into the objective function to obtain the objective function of New-SGNMF.The detailed derivation process of New-SGNMF update iteration rule is shown,and its convergence is proved.The experimental results show that New-SGNMF can achieve the desired effect on COIL20,PIE-pose09 and YaleB.This dissertation explores different image recognition approaches based on GNMF,and designs new approaches using the existing theories and technologies.The proposed method can filter the interference information in the process of image recognition through threshold optimization,and further sparse the basis matrix using norm constraint,so as to achieve the goal of accurate extraction of image features and improve the effect of image recognition. |