Nonconvex Low Rank Representation And Its Application In Image Clustering

In recent years,image clustering is a hot topic in machine vision and image processing.This is a worth studing thing that how to get effective image data quickly and accurately.With the rapid development of computer technology,the image is more valuable when it is converted into useful information.However,in practical applications,the image has complex features including large quantities,complex noise and uncertain structures.The clustering algorithm has been existed that cannot handle high dimensional image data effectively.Meanwhile,when the data contains a special structure,the clustering algorithm has some defects.For different image data structures and types,how to design a clustering algorithm with high accuracy and fast running speed that is an important goal and topic in machine vision field.Benefiting from theoretical advantages such as linear representation and low rank constraint,low rank representation and sparse methods have achieved satisfactory results in subspace clustering applications.However,due to consistency considerations of input features and suboptimal approximation of the constraint function which reduce the performance of clustering algorithm.With the problems of current research,this paper proposes two strategies which is used to improve representation performance of data.Firstly,incorporate the feature evaluation mechanism into the representation error of the sample,that is different from of the traditional low rank structure cumputes all input features simultaneously.This mechanism evaluates the contribution of different features to the representation matrix simultaneously.This operation is good for representing more accurate data relationships.Secondly,in order to estimate the singular values more accurately than the nuclear norm,this thesis propose to use a particular class of parameterize non-convex penalty functions.We show to set the nonconvex penalty rameter to ensure the proposed objective function is strictly convex.In this thesis,we use feature evaluation mechanisms to select features.,construct a coefficient matrix using data self-representation and propose to use a particular class of parameterized non-convex penalty functions.We propose a new model,namely subspace clustering via joint Low-rank Reinforcement and Feature Evaluation(LRFE).We derive an Alternating variable Multiplier Method to solve the proposed problem.Comprehensive experiments are conducted on synthetic data and image clustering to achieve excellent performance.In order to effectively solve the problem that the data in the actual application contains complex noise distribution,we constrain the noise error term using a weighting matrix that allows adaptive assignment of weighting factors.So,we can get a more accurate representation matrix.We further propose a algorithm namely subspace clustering via joint Iterative Reconstrained Low-rank Representation and Weighted Nonconvex Regularizer(LRWNR).Accelerated proximal gradient method is applied to solving the proposed algorithm.The experimental results show that the proposed algorithm has excellent clustering accuracy and operating efficiency under complex noise distribution.