An evolutionary algorithm with 2-D encoding for image segmentation

This paper presents an evolutionary approach which treats the image segmentation as a graph partitioning problem. An image is described as a weighted undirected graph where pixels correspond to nodes, and those pixels with similar values and positions are connected by edges. The weighted normalized cut criterion (WNcut) is used in this paper for this graph partitioning problem to measures both the dissimilarity between different partitions and the total similarity within the groups. This paper adopts a 2-dimensional representation of chromosome to directly present an image segmentation which is beneficial both to the genetic operators in the evolutionary process and to efficiently reduce the running time. In addition, the proposed evolutionary algorithm uses prior user's preference information to control the segments of the image through a random walker approach to initialize population. Experimental results demonstrate that our proposed algorithm is able to efficiently handle segmentation cases that segments images into several partitions based on human visual perception. The statistical results of entropy-based evaluation also suggest that our approach could achieve a more accurate segmentation.