Neighbor number, valley seeking and clustering

摘要: This paper proposes a novel nonparametric clustering algorithm capable of identifying shape-free clusters. is based on estimation the normalized density derivative (NDD) and local convexity distribution function, both which are represented in very concise form terms neighbor numbers. We use NDD to measure dissimilarity between each pair observations neighborhood build connectivity graph. Combined with convexity, this similarity can detect minima (valleys) separate different major demonstrate that has close relationship single-linkage hierarchical be viewed as its extension. The performance tested synthetic real datasets. An example color image segmentation also given. Comparisons several representative existing algorithms show proposed method robustly identify clusters even when there complex configurations and/or large overlaps.