Entropy-Based Multiscale Image Segmentation with Edge Refinement

An image segmentation algorithm that utilizes the concept of entropy is described. An index called Gray-scale Image Entropy (GIE) is used to measure the amount of information contained in a gray-scale image with reference to the underlying true scene. A prototype segmented image is produced and its quality is evaluated in terms of the GIE value the larger the GIE value the better would be the prototype segmented image. A multiscale entropy-based image segmentation algorithm is proposed to produce the prototype segmented images in successively refined spatial resolution until the best segmented image (in an information-theoretic sense) is found. The algorithm starts with detecting the best segmented image in a coarse spatial scale. The scale is then made finer and the object boundary pixels in that finer scale are detected and their classification status updated to produce a more refined segmented image with larger GIE value. The spatial scale is refined successively in a recursive manner until the best segmented image is obtained in the pixel scale. Simulation results presented in this paper demonstrate the feasibility of the proposed algorithm.