An Online Multiscale Clustering Algorithm for Irregular Data Sets

Abstract

Clustering analysis has been widely applied to image segmentation, however, many of them cluster convex data sets with an offline way. This paper proposes a new online multiscale competitive learning approach for irregular data sets. This approach defines the multiscale learning rule for data sets using local scattering information of data and starts to cluster from only one prototype. When new vectors in another cluster are input, the initial prototype self-splits and produces new prototype to represent them. The process continues until there is no more vector. The clustering scale is defined by the local variances of clusters. We carried out experiments on irregular data sets and obtained better clustering results compared to K-means algorithm.