Image Segmentation based on Random Process

Abstract

This thesis aims to advance research work in image segmentation by developing robust techniques for image segmentation. Segmentation of images is a major task of image processing. There is no general segmentation procedure that can deal with all sorts of images and provide good segmentation results for all of them. Also, the correct solution will always to a certain degree depend on subjectivity. Segmentation by thresholding is one of the most simple and fast method of segmentation. This thesis aims to provide research insights into field of segmentation of images based on image statistics. We concentrate on using the thresholding approach for segmentation in this thesis. This approach is utilized widely due to its simplicity of implementation and less computation cost. This approach uses histogram of image as a basic tool. This histogram is used to calculate threshold by various method. Initially we used an iterative algorithm for making clusters, grouping same type of pixels together and getting a threshold. We also discuss Otsu's algorithm and implement it on various images. Otsu algorithm is one of the best method of thresholding. It is fast and simple, working on 1D date only. Then we discuss segmentation using entropy of the image. Initially we did single level thresholding then we extended it to any arbitrary no of threshold as specified by user. We then propose a new algorithm which uses statistics of the image like mean, variance etc. to extract a feature vector image from the input image. It uses local properties to extract the feature vector which is then used for calculating thresholding using between class variance of the clusters formed.