Divisive Hierarchical K-Means

Abstract

This paper focuses on clustering methods for content-based image retrieval CBIR. Hierarchical clustering methods are a way to investigate grouping in data, simultaneously over a variety of scales, by creating a cluster tree. Traditionally, these methods group the objects into a binary hierarchical cluster tree. Our main contribution is the proposal of a new divisive hierarchy that is based on the construction of a non-binary tree. Each node can have more than two divisive clusters by detecting a better grouping in m classes (m??[2,5]). To determine how to divide the nodes in the hierarchical tree into clusters nodes, we use K-means clustering, [1]. At each node, to determine the correct number of clusters, we use a quality criterion called Silhouette. The solution that kmeans reaches often depends on the starting centroids, however we tested three methods of initialization, and we used the most suitable for our case.