A Dual Branch Segmentation Network for Forward-looking Sonar Image

Accurate segmentation of marine objects holds crucial significance and value in underwater resource exploration and marine rescue. However, in complex and noisy underwater environments, existing CNN-based algorithms face limitations such as insufficient global feature extraction capabilities and incomplete segmentation of certain targets in the sonar image. To overcome these inevitable problems, this article innovatively introduces a segmentation network based on transformer blocks. The algorithm encompasses a dual-branch structure that involves local and global feature extraction. The main components of the algorithm include four parts: Image preprocessing, Local feature extraction branch, Global feature extraction branch and Segmentation head. Finally, we conducted experiments on the public dataset, and the mIoU value of the proposed method surpasses other CNN-based algorithms. The results illustrate the effectiveness and feasibility of the segmentation network.