Sonar Image Segmentation Based on Markov Gauss-Rayleigh Mixture Model

In the process of the sonar image analysis, the information of the shadow and echo are very significant. However, most of the segmentation methods are failed to abstract the useful information from the background with strong noises. To overcome this problem, we apply the Markov random field to segment the features of sonar images, which expresses the spatial correlation of the image pixels sufficiently. A Gauss mixture model is proposed to fit the observation probability distribution given each class using vector quantization with a minimum discrimination information distortion. Based on our experiments, this method is more appropriate than traditional segmentation ones. Meanwhile, comparing with other segmentation methods, it can reduce the complexity and improve the segmentation efficiency.