A comparative analysis of entropy based segmentation with Otsu method for gray and color images

Image segmentation quality is usually governed by two main parameters associated with a particular segmentation method: threshold selection and seed-point selection. Various methods such as the histogram method, entropy based method, busyness measure methods, etc. are well known for threshold selection in the image segmentation problems. In this thesis, threshold selection is done on the basis of different entropy measures on both grayscale and color images. Comparative study of the Shannon and non-Shannon entropies (Renyi, Havrda-Charvat, Kapur and Vajda) with Otsu Method is done to obtain image segmentation. We classify these methods according to their evaluation criteria viz. PRI (Probabilistic Rand Index), VOI (Variation of Information) and GCE (Global Consistency Error). These underlying metrics and combination methods help in determining the performance of an evaluation measure. It is observed through the simulation experiments performed on images, that the position of the smallest minima obtained in the entropy versus gray-level plot is different for each entropy measure. For a particular definition of the entropy plots are generated and the threshold values obtained from these plots possess the segmentation results. It is observed that Havrda-Charvat entropy measure is better matched with Otsu Method than other entropy measures.