Sonar image segmentation using the angular dependence of backscattering distributions

In high resolution sonar imaging, the K-distribution has shown interesting properties for describing the statistics of backscattered intensity. This distribution is a function of two parameters. The first one is the mean intensity which is classically angular dependent. It has been found that the second one, the shape parameter, is also angular dependent. The K-distribution shape parameter shows two types of angular dependence. First, for rough sea floors, the parameter increases monotonously from small incident angles to grazing angles. And for a smooth seabed, after an identical increase, the shape parameter decreases for most grazing angles. In this paper, the angular dependence is first used to improve sonar images segmentation and then a new segmentation algorithm is proposed combining a probabilistic SVM-based classification outputs and a Markov-based regularization criterion. This was tested on sonar images provided by the SIMRAD EM1000 multibeam echosounder (MBES).