Sparseness and Correntropy-Based Block Diagonal Representation for Robust Subspace Clustering

Yesong Xu, Ping Hu, Jiashu Dai, Nan Yan, Jun Wang

Abstract: Block diagonal representation, which aims to compel the desired representation coefficient to have a block diagonal structure directly, has extensive applications in the domains of computer vision and machine learning. However, single residual modeling in existing works is not robust enough when handling complex noise (i.e., sparse noise and impulsive noise) in reality. To overcome this challenge, a novel Sparseness and Correntropy-based Block Diagonal Representation (SC-BDR) model is proposed, which is able to pursue ideal block diagonal representation and effectively deal with various types of noise. Furthermore, the corresponding optimization algorithm is designed for the proposed problem, and we also conduct extensive experiments to demonstrate the robustness and effectiveness of the SC-BDR model on real-world data.