Abstract: Vision problem ranging from image clustering to motion segmentation can naturally be framed as subspace segmentation problem, in which one aims to recover multiple low dimensional subspaces from noisy and corrupted input data. Low rank representation-based subspace segmentation algorithm (LRR) formulates the problem as a convex optimization and achieves impressive results. However, it needs to take a long time to solve the convex problem, and the clustering accuracy is not high enough. Therefore, this paper proposes a distributed low rank representation-based sparse subspace clustering algorithm (DLRRS). DLRRS adopts the distributed parallel computing to get the coefficient matrix, then take the absolute value of each element of the coefficient matrix, and retain the k largest coefficients per column and set the other elements to 0 to get a new coefficient matrix. Finally, DLRRS performs spectral clustering over the new coefficient matrix. But it doesn’t have incremental learning function, so there is a scalable distributed low rank representation-based sparse subspace clustering algorithm (SDLRRS) here. If new samples are brought in, SDLRRS can use the former clustering result to classify the new samples to get the final result. Experimental results on AR and Extended Yale B datasets show that the improved algorithms can not only obviously reduce the running time, but also achieve higher accuracy, which verifies that the proposed algorithms are efficient and feasible.

Key words: low rank representation, subspace clustering, parallel computing, incremental learning, coefficients reconstruction