Unsupervised segmentation of underwater objects in sonar images

We focus on the segmentation of sonar images for the aim of underwater object detection. Speckle noise and intensity inhomogeneity may cause false detections, and complex seabed textures like sand-ripples and sea-grass often lead to false segmentation. To tackle these problems, we propose a new method to incorporate the possible spatial correlation between neighboring pixels in the sonar image for improved segmentation. Our method modifies the expectation-maximization (EM) algorithm by adding an intermediate step (I-step) between the expectation (E-step) and maximization (M-step). Results show that our proposed method achieves improved segmentation performance over the state-of-the-art and is also robust to different seabed texture and for intensity inhomogeneity.