Mine classification based on raw sonar data: an approach combining Fourier descriptors, statistical models and genetic algorithms

In the context of mine warfare, detected mines can be classified from their cast shadow. A standard solution is to perform image segmentation first (we obtain binary from graylevel image giving the label zero for pixels belonging to the shadow and the label one elsewhere), and then to perform a classification based on features extracted from the 2D-shape of the segmented shadow. Consequently, if a mistake happens during the process, it will be propagated through the following steps. In this paper, to avoid such drawbacks, we propose a novel approach where a dynamic segmentation scheme is fully classification-oriented. Actually, classification is performed directly from the raw image data. The approach is based on the combination of deformable models, genetic algorithms, and statistical image models.