Subspace Clustering and Feature Extraction Based on Latent Sparse Low-Rank Representation

Robust recovery of multiple subspace structures from high-dimensional data with noise has received considerable attention in computer vision and pattern recognition. Low-Rank Representation (LRR) as a typical method has made satisfactory results in subspace clustering. Latent Low-Rank Representation (LLRR) is an advanced version of LRR, which considers the row and column of data to solve the insufficient samples problem. However, they fail to exploit the local structures of data. To address this problem, Latent Sparse Low-Rank Representation (LSLRR) is proposed to capture the local and global structures of data by considering sparse and low-rank constraints simultaneously. In this way, LSLRR not only solves the clustering problem, but also extracts significant features for classification. Inexact Augmented Lagrange Multiplier method (IALM) is utilized to solve its objective function. Experimental results in subspace clustering and salient features extraction demonstrate the proposed LSLRR have a favorable performance.