EHDS-YOLO: a lightweight side-scan sonar object detection model for intelligent underwater inspection in marine ranching

Abstract

Accurate underwater target detection is essential for intelligent inspection, facility monitoring, and environmental perception in marine ranching. However, side-scan sonar images are often affected by weak target echoes, blurred boundaries, seabed reverberation, and speckle noise, which limit the performance of conventional detection models in complex aquaculture environments. To address these challenges, this study proposes EHDS-YOLO, a lightweight YOLO11-based detector for marine-ranching inspection. The proposed framework integrates sonar-oriented feature enhancement, lightweight feature fusion optimization, and a compression–recovery strategy combining pruning and knowledge distillation to improve detection efficiency while preserving accuracy. Furthermore, LAMP pruning and BCKD knowledge distillation are employed to compress the model and recover pruning-induced performance degradation. A side-scan sonar dataset, PPRN-SA, was constructed from field sea trials in the Yellow Sea and includes four typical marine ranching targets: aquaculture cages, pipelines, reefs, and bridge piers. Experimental results show that EHDS-YOLO achieves an mAP50 of 75.7% with 1.27 M parameters and 4.4 GFLOPs. On the Jetson Orin Nano Super, it reduces single-image processing latency from 34 to 20 ms and increases throughput from 29 to 50 FPS relative to YOLO11n. After independent retraining on SSS-Mine, EHDS-YOLO achieves an mAP50 of 82.7% and an mAP50:95 of 51.6%. These results support its use as a compact deployment-oriented detector, although fine-grained localization remains a limitation.