Research On Robust Subspace Clustering Algorithm And Its Application

With the widespread application of emerging information technologies such as cloud computing,artificial intelligence,and the Internet of Things in important fields such as industrial manufacturing,transportation,mobile telecommunications,e-commerce,biomedicine,and aerospace,high-dimensional data continues to emerge.High-dimensional data not only provides people with richer and more comprehensive information,but also increases the difficulty of data processing.Especially when processing visual data such as face images and videos,these data not only have high-dimensional features,but also often hide a large amount of fragmented information,error information,and noise,making the data exhibit complex nonlinear characteristics.The subspace clustering algorithm represents high-dimensional data as a linear combination of multiple low-dimensional subspaces,which is widely regarded as an effective way to process high-dimensional data.However,traditional subspace clustering algorithms have difficulty capturing the nonlinear structure in the data.The subspace clustering algorithm based on kernel mapping can use kernel tricks to achieve linear representation of nonlinear data in Rendering Kernel Hilbert Space(RKHS),but noise will destroy the subspace structure,making the design of kernel function become difficult and even cause essential information to be obscured.This paper focuses on how to improve the robustness of the subspace clustering method in processing low-quality visual data clustering such as face images and videos.In view of the existing problems of damaged subspace structure,difficulty in selecting kernel functions and data occlusion,research is conducted in terms of consistent representation of data structure relationship,multi-kernel automatic weighted learning,and adaptive data correction.The main research contents and results of this paper are as follows:1.Research on robust subspace clustering method based on consistent representation of data structure relationships.In order to address the problems of structural inconsistency between the original space and the kernel mapping subspace in the process of nonlinear data subspace clustering by the kernel mapping subspace clustering method,as well as the noise interference problem in nonlinear data,a Rank-Constrained Block Diagonal Subspace Clustering model and its extension models,along with their respective solving algorithms,are proposed.These models combine the low rank of the mapped data and the block diagonal regularization constraint of the representation coefficient matrix,and use the Correntropyinduced metric theory to design the Correntropy-induced metric method based on kernel mapping.The experimental results demonstrate that the proposed model and their extension models achieve lower error segmentation rates and exhibit strong robustness.2.Research on robust subspace clustering method based on multi-kernel automatic weighted learning.In addition to being interfered by data corruption,multi-kernel subspace clustering methods also face the issue of kernel noise caused by the lack of reliable fundamental kernel combination criterion in multi-kernel learning.To address the mixed noise problem in multi-kernel subspace clustering,this thesis utilizes the kernel order distance method to design a multi-kernel weight automatic allocation strategy and an enhanced sparse noise reconstruction method.Based on these studies,the Enhanced Robust Subspace Clustering Method with Multi-Kernel Automatic Weighting Learning(EMAWSC)model and its corresponding solving algorithm are proposed.The experimental results demonstrate that EMAWSC achieves significant clustering accuracy and robustness.3.Research on robust graph representation clustering method based on adaptive data correction.Aiming at the problem of how to improve the robustness of subspace clustering methods in processing occluded low-quality visual data clustering,an Adaptive Data Correction-Based Graph Clustering(ADCGC)model and corresponding solving algorithm are proposed.This model utilizes the robust principal component analysis to design a method for reducing intra-class variation,and incorporates graph learning theory.The experimental results indicate that ADCGC enhances the robustness when dealing with low-quality singlechannel visual data tasks.Meanwhile,to solve the clustering problem of low-quality multichannel data such as color face images and videos,this thesis explores multi-kernel robust principal component analysis,multi-channel technology,consensus graph learning and data imputation technology,and subsequently proposes an Adaptive Data Completion-Based Multiple Kernel and Multiple Graph(ADCMKG)model and its corresponding solving algorithm.The experimental results on various real datasets demonstrate that ADCMKG have better robustness and generalization capabilities.This paper aims at addressing the challenges of subspace clustering for low-quality visual data.It conducts research in the areas of consistent representation of data structural relationships,multi-kernel automatic weighted learning,and adaptive data correction to tackle issues such as compromised subspace structure,difficult kernel function selection,and data occlusion.The experimental results on various diverse datasets demonstrate the effectiveness of the proposed methods.Particularly,ADCGC is applied to some low-quality single-channel data processing tasks,such as face image recognition,disguised face restoration,image restoration,and video background recovery,it shows significant advantages.Simultaneously,ADCMKG has also achieved excellent performance in tasks such as face recognition of low-quality multi-channel images,object segmentation of lowquality multi-channel videos,and object segmentation of low-quality invisible light videos.