Recent review on image clustering

Abstract

In this review, image clustering problem is discussed starting from global learning based clustering approaches such as Kmeans to the recent challenges in this domain. In global learning based clustering models, cluster evaluation criteria was formulated on whole image datasets. In local learing based clustering models, local neighbourhood information in image data matrices were utilised. However, performances of local learning based clustering models may face limitations for image datasets that contain images with pose, illumination, or occlusion variations. Further, both global learning and local learning approaches were combined in clustering models. Still, image clustering performances are not significantly improved. I elaborate challenges in image clustering problem by categorising image datasets as Gaussian-like or multimodal. Clustering performances on 14 benchmark image datasets of almost all state-of-the-art clustering models are optimal only for Gaussian-like image datasets. Thus, image clustering performance has direct correlation with the distribution of image datasets. Further, by employing optimal image descriptor, clustering performances are data dependent. Almost all exiting clustering models are based on second-order statistics. Owing to which, multimode image patterns may not be effectively handled even by exploiting both local and global information in image data matrices.