Image segmentation by utilizing entropy concept

Abstract

In this thesis, the template matching approach is employed in image segmentation where the search for the best template is guided by an entropy criterion. The Grayscale Image Entropy (GIE) is used to evaluate the goodness of a given template against a grayscale image. Given a grayscale image, a template is postulated to approximate the true scene that gives rise to the grayscale image. For the grayscale image and the template, the GIE is calculated. It is shown that the greater the GIE value, the better would the template resemble the true scene. The template is then adjusted until the maximum GIE value is resulted. The template is the optimum segmented image in an entropy sense. Based on the template matching approach and the maximum GIE criterion, three new image segmentation algorithms are proposed and investigated in this thesis. The first algorithm detects the boundary of the object in an image by evaluating the GIE value. It is shown that when the object in a template is at a position overlapping the true scene object boundary, the resultant GIE will be zero. By connecting all these zero GIE value points, the boundary of the true scene object can be detected and then the grayscale image segmented. The second and the third segmentation algorithms start with an initial template. Then the classification status of the template pixels adjusted either on a pixel-by-pixel or a block-by-block basis. Each adjustment would result in a new template and a new GIE value. Only those changes that lead to greater GIE value will be retained. After a series of adjustments made to the template, the optimum template that gives rise to the maximum GIE is obtained as the optimum segmented image. Segmentation results for synthetic and real images obtained in this thesis justify the approach.