Research On Image Feature Expression And Classification Algorithm Based On Non-negative Matrix Factorization

With the increasing development of big data,especially the rise of intelligent data,image data not only become a massive deep data,but also further into the high-dimensional wide data.How quickly and efficiently retrieve,classify and extract valuable information from these massive and high-dimensional large-scale image data has become a hotspot in the research of modern science and technology society.However,image feature representation and classification,as an important part of machine learning and data mining,has been widely applied in many fields,such as image processing,data analysis,market research and so on.In the existing matrix decomposition technique,non-negative matrix factorization(NMF)as an effective method of data representation has received extensive attention and deeply studied.Therefore,the exploration of NMF has the certain practical significance.On the basis of NMF,this paper mainly studies how to apply various constraints to NMF framework,how to effectively fuse features of NMF and how to combine NMF with image classification better.According to the content of above research,three improved NMF methods are proposed:(1)graph regularized and incremental NMF with sparseness constraints.The algorithm not only preserves the geometric structure,but also makes full use of the results of previous step to learn incrementally and the sparsity limit is applied to the coefficient matrix.Finally it is integrated into the same objective function,which can reduce the operation time and has a better classification accuracy;(2)multi-constrains NMF based on feature fusion.On the basis of various constraints,the algorithm not only considers the class information and sparse constraints,but also introduces the graph regularization,and fuses the decomposed image features with different sparsities to enhance the clustering performance and effectiveness;(3)dual graph-regularized constrained NMF.The algorithm not only considers the geometric structures of data manifold and feature manifold simultaneously,but also the known label information is restricted to the NMF,which improves the quality of learning and enhance the recognition performance.This paper verifies the convergence,sparsity and identification of improved aigorithms,and proves the feasibility of which on common benchmarks.According to the experimental results,it can be further analyzed the effectiveness of three methods.Finally,the clustering effects of three algorithms are compared,and the results shown that the DCNMF is the best.