Natural and Remote Sensing Image Segmentation Using Memetic Computing

Abstract

In order to solve the image segmentation problem which assigns a label to every pixel in an image such that pixels with the same label share certain visual characteristics more effectively, a novel approach based on memetic algorithm (MISA) is proposed. Watershed segmentation is applied to segment original images into non-overlap small regions before performing the portioning process by MISA. MISA adopts a straightforward representation method to find the optimal combination of watershed regions under the criteria of interclass variance in feature space. After implementing cluster-based crossover and mutation, an individual learning procedure moves exocentric regions in current cluster to the one they should belong to according to the distance between these regions and cluster centers in feature space. In order to evaluate the new algorithm, six texture images, three remote sensing images and three natural images are employed in experiments. The experimental results show that MISA outperforms its genetic version, the Fuzzy c-means algorithm, and K-means algorithm in partitioning most of the test problems, and is an effective approach when compared with two state-ofthe-art image segmentation algorithms including an efficient graph-based algorithm and a spectral clustering ensemble-based algorithm.