Research on segmentation of sonar image based on features learning

A side-scan sonar image often contains a lot of noises and its resolution is very low, the edge of side-scan sonar image is blurred, so these characteristics make it difficult to segment objects. The current common segmentation algorithms generally require setting parameters manually, and these parameters are closely related with the images they collected. Consequently, it is difficult to achieve universality and autonomy. According to the properties of segmentation algorithm and characteristics of side-scan sonar image, this paper proposes an algorithm based on unsupervised feature learning. We use sparse auto-encoder to learn the local binary patterns and Haar-like feature of the side-scan sonar images, then we build an exclusive feature for them. In the end, we test this method by using clustering algorithm based on features we built previously to segment several side-scan sonar images. The results show that the algorithm we proposed has better performance in both adaptability and effectiveness.