Spectral clustering based on dictionary learning sampling for image segmentation

Abstract

The classical spectral clustering spends much time on calculating similarity matrix and vectors. Nyström can be used to obtain approximate similarity matrix, which clustering results are unstable since randomly selected the sampling points. And dictionary learning is used to search the example points to construct similarity matrix, which Dictionary learning sampling Nyström spectral clustering algorithm (DNSC) is proposed in this paper. The DNSC constructs the dictionary by learning the entire data set, and select those points nearest atoms of the dictionary as sampling points. Further, removing redundancy points from candidate sampling points set that makes the size of sampling points be adaptive and sufficient. Compared with the Nyström Spectral Clustering Algorithm and K-means Nyström Spectral Clustering Algorithm, the presented algorithm has more stable and more accurate segmentation results. © 2012 Springer-Verlag.