Model driven image segmentation using a genetic algorithm for structured data

Abstract

In this paper, a method, integrating efficiently a semantic approach into an image segmentation process, is proposed. A graph based representation is exploited to carry out this knowledge integration. Firstly, a watershed segmentation is roughly performed. From this raw partition into regions an adjacency graph is extracted. A model transformation turns this syntaxic structure into a semantic model. Then the consistence of the computer-generated model is compared to the user-defined model. A genetic algorithm optimizes the region merging mechanism to fit the ground-truth model. The efficiency of our system is assessed on real images. © 2010 Springer-Verlag.