Evaluation of the number of segments using weighted entropy

Image segmentation plays a very important role in many image understanding and computer vision techniques. Over the last few decades image segmentation techniques have been advanced significantly, though a few quantitative evaluation techniques have been proposed. Most of the segmentation evaluation techniques are largely subjective that means the evaluation is done visually, intuitively or qualitatively from the comparison of several segmented images. Sometimes the evaluation is even done based on the results of the segmentation step in the subsequent application or processing step. A novel objective segmentation evaluation method based on entropy is proposed in this paper which is simple to implement. In this proposed method the weighted self-entropy and mutual-entropy are used to measure the dissimilarity of the pixel characteristics among the segmented regions and the similarity within a region. This proposed evaluation method produces a numeric score for a segmented image and thus can be used for the comparison of different segmentation algorithms of an image. Moreover, we can use this method to find the values of the parameters of any segmentation algorithm that best suit for a given image. The simulation results show that the presented segmentation evaluation method is effective in identifying both over- and under-segmentation, and the ”good” segmentations that match well with the reference. The preliminary outcomes also open up new areas for further research on objective segmentation evaluation using entropy.