Adaptive clustering based segmentation for image classification

Image segmentation based on clustering low-level image features such as colour and texture, has been successfully employed in image classification and content-based image retrieval. In segmentation based image classification, the role of clustering to segment an image into its relevant constituents that represent image visual content as well as its semantic content. However, image content can vary from having a simple foreground object on a regular background to having multiple objects of different sizes, shapes, colour and texture in complex background scenes. This makes automatic image classification a challenging task. This paper evaluates three adaptive clustering algorithms of different categories, i.e., partition-based, model-based, and density-based in segmenting local colour and texture features for image classification. Experiments are conducted on the publicly available WANG database. The results show that the adaptive EM/GMM algorithm outperforms the adaptive k-means and mean shift algorithms.