Local consistent low rank representation for image clustering

Abstract

Low rank representation (LRR) is one of the state-of-the-art methods for subspace clustering, which has been widely used in machine learning, data mining, and pattern recognition. The main objective of LRR is to seek the lowest rank representations for the data points based on a given dictionary. However, there are some drawbacks in the current LRR-based approaches: 1) the original data usually contain noise and may not be representative as a dictionary; 2) only global Euclidean structure is considered, while the local manifold structure, which is often important for many real applications, has been ignored. To this end, we propose an improved LRR-based approach, called Local Consistent Low Rank Representation (LCLRR), in which, the dictionary learning and low rank representation can be achieved simultaneously, and both the global Euclidean structure and the local manifold structure of original data can be exploited. An efficient optimization procedure, which is based on Alternating Direction Method of Multipliers (ADMM), is used for LCLRR optimization.