Target Detection in Sidescan Sonar Images via Improved YOLOv8

ABSTRACT

This paper proposes an improved YOLOv8 model integrating LSKA, BoT3 and Shape‐IoU loss. LSKA strengthens multi‐scale feature extraction; BoT3 is selectively embedded in P3 and P4 layers to model long‐range dependencies and suppress acoustic‐shadow‐induced false alarms; Shape‐IoU optimises box regression for irregular wreck geometries. Experiments on our self‐built sonar dataset achieve 0.886 precision and 0.893 mAP@0.5. Compared with vanilla YOLOv8, our model reduces parameters from 11.1 to 10.3 M and GFLOPs from 28.4 to 26.2, at the cost of moderate FPS decline. Ablations validate the complementary effects of modules and the optimal BoT3 embedding scheme. Visual results verify improved target‐shadow separation, and cross‐dataset evaluations on SCTD and LSOD confirm strong generalisation. This method offers an effective solution for underwater wreck detection in complex sidescan sonar scenarios.