A novel method for image clustering

Abstract

Image clustering has been attracting mounting focus on widely used fields, such as data compression, information retrieval, character recognition and so on, due to emerging applications of various web-based and mobile-based image re- trieval and services. To study this, based on Voronoi diagram, we propose a novel image clustering algorithm to effective discovery of image clusters in this paper. More specifically, based on Voronoi diagrams at first, a number of irregular grids are built across whole plane. Furthermore, leveraging good property of the nearest neighbor for Voronoi diagrams, various irregular grids of plane are assigned by points to different clusters. On one hand, based on density of grid points, it automatically adjusts final suitable number of clustering; on other hand, according to changes of centroids, it tunes positions for Voronoi's seeds. At last, Voronoi cells finally become result of clustering process. The empirical experiment results show that our proposed method not only can cluster image dataset effectively, but also can achieve comparative performance with X-means algorithm and K-means algorithm. Moreover, our proposed method can outperform effectiveness for both DBSCAN and OPTICS algorithms, which are classic density-based clustering algorithms towards larger- scale real-world applications.