Low-Rank Representation And Clustering Of Image Data

With the rapid development of the science and technology,most information has been presented in the forms of digital pictures in recent years.Moreover,it has become much easier for people to obtain these data.At the same time,as increasing of the scale of data set,the storage,transmission and processing of the images are often beyond the human's own ability.Thus more and more academics and industrial researchers are concerned about seeking effective image analysis and relevant processing methods.In terms of image data processing,a single image is usually vectorized,and a plurality of vectorized images consist of a sample matrix.In the matrix,the samples of the same class are highly correlated while those of the different classes are not.Such sample matrix is potentially a low-rank matrix.Based on the low-rank property of the image data,three models are proposed in this paper for high dimensionality and easy incompletion of feature.The main tasks are as follows:1.In this paper,the adjustable arc tangent latent low-rank representation subspace clustering algorithm is proposed based on the low-rank characteristics of image.The traditional low-rank representation subspace clustering algorithm is according to the idea of self-representation of samples,and low-rank constraint is applied to the representation coefficients matrix to get low-rank representation of the sample data.The direct use of rank function constrain on the coefficient matrix is a NP-hard problem.The common method is to replace the rank function with nuclear norm.Nevertheless,the nuclear norm sums the nonzero singular of the matrix and it can be influenced by the singular value of the matrix.Therefore,In this paper,the rank function is replaced by the adjustable arc tangent function.Experimental results on several image data sets show that this method can effectively improve the clustering accuracy of the image data.2.Aiming at the problem that the image is easily incomplete,the matrix completion low-rank subspace clustering algorithm is proposed.Being covered and damaged,the images are prone to feature deletion,which directly affect the clustering result.The general method is to f-irst complete the data,and then do further analysis,which ignores the relevance between the data completion and clustering processing.This paper integrates the completion and clustering of image data in order to solve effectively with the clustering problem for the incomplete image data.3.Unsupervised feature extraction algorithm based on the low rank sparse similarity preserving is proposed to solve the high dimensionality problem of image data.The existing image processing algorithms keep the local structure of sample before and after the dimension reduction,and ignore the relations among all the samples.The low rank of the matrix characterizes the global structure of the whole samples,and it is incorporated to the existing feature extraction method,which can effectively overcome the shortcomings of the existing method that only consider the local structure.The simulation results in this paper show that the method can effectively extract the feature of image data.