Auto-segmentation using mean-shift and entropy analysis

In this paper we discuss the automization of the mean-shift clustering algorithm which is dependent on user-defined parameters for its efficiency. We propose that the optimum solution of mean-shift corresponds to the maximum entropy of the mode-probabilities which assume the form of a uniform probability distribution for the meaningful segmentation of a scene. The experimentation on the benchmark Berkeley segmentation database shows high accuracy for the automated mean-shift as compared to the baseline method. The Non-extensive entropy with Gaussian gain gives highly meaningful segmentation that agrees with human perception as compared to the extensive Shannon entropy.