Segmentation markovienne d'images sonar haute résolution

The study of the high resolution sonar image segmentation using a Markov model of the classes eld lead us to propose a new modelisation for the noise. This last, by introducing the markovianity on the observation elds knowing the classes, allows us to take into account the correlation without losing the fact that the marginal statistics are not gaussian (K law or gamma law). To check the contribution of this aproach, we vary the correlation between the pixels of synthetic images before restoring them by the MPM procedure in a supervised case. We observed then some amelioration of the segmentation for certain values of the correlation coe cient as compared to the usual case where the noises are independent. 1 Contexte et pr esentation