CIA-SOLO: Coordinate Instance Attention SOLOv2 for Instance Segmentation in Side-Scan Sonar Images

Abstract

As the main sensor for underwater target detection, side-scan sonar plays an important role in seabed topographic survey, shipwreck salvage and other tasks. However, affected by the reverberation noise on the seabed, the side-scan sonar image contains a large amount of speckle noise, which makes the target boundary information weak and the segmentation accuracy low. In addition, targets in side-scan sonar images are often accompanied by shadows, and the locations and shapes of these shadows are similar to those of the targets, which seriously affects the accuracy of instance segmentation. In order to solve these problems, we propose an improved SOLOv2 algorithm based on coordinate attention module and instance attention module (CIASOLO). Specifically, after the backbone of SOLOv2, coordinate attention is used to determine the regions of interest (ROI) in the feature maps, and then the specific attention maps are learned through the attention module, which makes the target and shadow more distinguishable. We have done a lot of experiments on Sonar Common Target Detection (SCTD) dataset, and the experimental results show that CIA-SOLO achieves 80.6% mAP@0.5, and is 7.3% higher than native SOLOv2.