Unsupervised Hierarchical Markovian Segmentation Of Sonar Images

This paper is concerned with hierarchical Markov Random Field (MRF) models and with their application to sonar image segmentation. We present a novel unsupervised hierarchical MRF model involving a pyramidal label eld and a scale-causal and spatial neighborhood structure. This allows us to more precisely model the local and global characteristics of image content for diierent scales. Such connections lead to eeciently propagate interactions and this approach seems to be well suited for the segmentation of very noisy sonar images. The MRF prior model parameters are estimated in an accurate and fast way simultaneously to the segmentation process by generalizing the estimation method proposed by Derin et al. Experiments with real images indicate that the proposed SCM algorithm (Scale Causal Multigrid) performs better than other hierarchical schemes for sonar image segmentation.