Research On Subspace Clustering Algorithm Based On Low Rank

In recent years,with the development of big data,the access to data has become broader and the requirements for data processing have become more stringent.In the field of computer vision,with the explosive growth of high-dimensional datasets,existing clustering algorithms are limited by the complexity of space and the accuracy requirements cannot be met.Due to the nature of subspace mapping,subspace clustering algorithms have attracted more and more attention.The spectral clustering algorithm combined with graph theory is an important branch of the subspace clustering algorithm.It essentially transforms the data matrix into an undirected graph.After the cut set criteria,the sub-graph completes the clustering task.Since the generated matrix directly affects the accuracy of the clustering,it becomes especially important how to process the data set to convert it into an ideal matrix.The existing generator matrix algorithm removes the noise construction matrix based on sparse ideas,but it has several disadvantages: One is that the data dictionary generated by using the data set will contain noise which belongs to data set and the data dictionary will be polluted;Second,the generated matrix will encode similar features to different coefficients.Third,the accuracy rate needs to be improved.Therefore,in the study of subspace clustering algorithms,how to design an efficient and accurate matrix generation algorithm is a problem worthy of study.This paper studies the subspace clustering algorithm and proposes two subspace clustering algorithms for specific problems:(1)In order to solve the problems of pollution data dictionary problems and accuracy,this paper proposes a sparsity-based low-rank subspace algorithm.The algorithm uses sparsity constraints to force the data in the data set to be linear representation of other data,complete the denoising of the original data set to get a clean dictionary,and then make the generated representation matrix have the ideal diagonal structure as possible.Compared with the existing algorithms, the experimental results show that the algorithm has good clustering accuracy.(2)This paper proposes a non-negative local constraint low-rank subspace clustering algorithm for the problem of unclean dictionary and generator matrix error coding.Based on the low rank constraint and the robust principal component analysis,the algorithm finds the best projection of the data in the low dimensional space and completes the denoising of the dictionary.At the same time,non-negative constraints are added to eliminate the problem of partial coding of similar features.Considering that the complexity of the model is increased after adding multiple constraints,a linear alternating direction adaptive method is introduced to iteratively solve the problem.From the experimental results of multiple experimental data sets,it is shown that the algorithm has significantly improved the accuracy of clustering compared to existing algorithms.