Dempster-Shafer Evidence Theory Supported EM Approach for Sonar Image Segmentation
Tai Fei, Dieter Kraus · 2017
In this paper, an expectation-maximization (EM) approach assisted by Dempster-Shafer evidence theory for image segmentation is presented. The images obtained by synthetic aperture sonar (SAS) systems are segmented into highlight, background and shadow regions for the extraction of geometrical features. Firstly, the proposed approach chooses the likelihood function given by Sanjay-Gopal et al. This likelihood function decouples the spatial correlation between pixels far away from each other. Secondly, the mostly implemented Gaussian mixture model is substituted by a generalized mixture model which adopts the Pearson system. As a consequence, the proposed approach is able to approximate the statistics of sonar images with more flexibility. Moreover, an intermediate step (I-step) is introduced between the E- and M-steps of the EM algorithm. The I-step adopts the Dempster-Shafer evidence theory based clustering technique to consider the spatial dependency among neighboring pixels. The states of neighbors are viewed as evidence to support the hypotheses about the state of the pixel of interest. Finally, the proposed approach is applied to SAS imagery to evaluate the performance.