Sonar image segmentation based on GMRF and level-set models

We propose two new level-set models to address the segmentation problem in sonar images. Local texture features, extracted using the Gauss-Markov random field model, are integrated into level-set energy functions to dynamically select regions of interest. Then, new two-phase level-set and multiphase level-set models are obtained by minimizing each new energy function, and the selection of model parameters is analyzed. The proposed models do not require re-initialization, which is usually a very costly procedure. Segmentation experiments on both synthetic and real sonar images show that the proposed two level-set models are accurate and robust when they are applied to noisy sonar images.