An Improved Structured Low-Rank Representation for Disjoint Subspace Segmentation

摘要

Low-rank representation (LRR) and its extensions have shown prominent performances in subspace segmentation tasks. Among these algorithms, structured-constrained low-rank representation (SCLRR) is proved to be superior to classical LRR because of its usage of structure information of data sets. Compared with LRR, in the objective function of SCLRR, an additional constraint term is added to compel the obtained coefficient matrices to reveal the subspace structures of data sets more precisely. However, it is very difficult to determine the best value for the corresponding parameter of the constraint term, and an improper value will decrease the performance of SCLRR sharply. For the sake of alleviating the problem in SCLRR, in this paper, we proposed an improved structured low-rank representation (ISLRR). Our proposed method introduces the structure information of data sets into the equality constraint term of LRR. Hence, ISLRR avoids the adjustment of the extra parameter. Experiments conducted on some benchmark databases showed that the proposed algorithm was superior to the related algorithms.