High quality segmentation of synthetic aperture sonar images using the min-cut/max-flow algorithm

In the context of automatic detection and classification for mine hunting applications, a high quality segmentation of sonar images is mandatory. Assuming a Markov Random Fields representation of the images, we propose a min-cut/max-flow segmentation algorithm. We introduce an original initialization of the graph cut algorithm based on the segmentation result of an Iterative Conditional Modes (ICM) segmentation approach. A large database of synthetic aperture sonar images containing 378 spherical and cylindrical man made objects has been segmented using both the ICM algorithm and the graph cut approach. Both sets of results have been automatically classified according to a set of significant features. Results are compared.