Image Retrieval Using a Hierarchy of Clusters

Abstract The goal of this paper is to describe an efficient procedure for color-based image retrieval. The proposed procedure consists of two stages. First, the image data set is hierarchically decomposed into disjoint subsets by applying an adaptation of the k-means clustering algorithm. Since Euclidean measure may not effectively reproduce human perception of a visual content, the adaptive algorithm uses a nonEuclidean similarity metric and clustroids as cluster prototypes. Second, the derived hierarchy is searched by a branch and bound method to facilitate rapid calculation of the k-nearest neighbors for retrieval in a ranked order. The proposed procedure has the advantage of handling high dimensional data, and dealing with non-Euclidean similarity metrics in order to explore the nature of the image feature vectors. The hierarchy also provides users with a tool for quick browsing