Memetic Algorithm For Image Segmentation Based On Normalized Cut

Image segmentation is a key researching direction of computer vision, pattern recognition and artificial intelligence, which is increasingly needed in applications of industry, military affairs, remote sensing, meteorology and robot vision. In fact, image segmentation is a NP hard optimization problem, whose aim is to label pixels for different clusters of the image. The result of optimization is to get a classification status in the visual for all pixels in the image. Unfortunately, it is difficult for NP hard problems to find the global optima in the limited space and time. Therefore, traditional methods usually simplify those NP hard problems to simpler optimization problems by prior information and mathematical methods, which gets better results. By analysing image segmentation methods existing, we try to solve this problem from another new way. Based on that evolutionary computation is often used to solve NP hard problems better, we present Memetic algorithm for image segmentation which cooperates evolutionary algorithm and image structure information.The researching content in this paper is as follows:1. Clonal selecting Memetic algorithm for image segmentation based on Normalized Cut is presented, because Memetic algorithm is a heuristic method based on population, which can evolute several solutions simultaneously to find the globally optima more efficient. We design the coding style and neighbourhood learning operator based on image structure to guide the searching direction in the evolution, which can accelerate the convergence. This method can stably converge to a global optima.2. On the basis of the first method, inverse neighbourhood learning operator and elitists preserving strategy are presented to accelerate the converging speed, which prove to be effective. And the terminating condition of circulating is improved.3. Directing at the high space complexity mentioned above, image level is added to the method in this paper. By prolonging images, the space complexity can be reduced and efficiency of the algorithm is improved.