A Lightweight Object Detection Model for ROV‐Based Underwater Foreign Object Salvage

ABSTRACT

Addressing challenges in underwater industrial maintenance, such as degraded quality, target occlusion and platform constraints, this paper proposes UGS‐YOLO_SP, a high‐precision lightweight model for foreign object recovery using a remotely operated vehicle (ROV). To mitigate data scarcity, a real‐world underwater operation dataset is first constructed. The study then introduces the gated strip‐aware spatial separation (GS3) block to enhance spatial context modelling for slender structures and the lightweight enhanced shared detail convolutional detection (LESCD) head to reduce redundancy via parameter sharing. This paper presents a self‐distillation and pruning (SP) lightweight framework that applies bridging cross‐task protocol inconsistency self‐knowledge distillation (BCSKD) for feature discrimination against underwater interference and layer‐adaptive magnitude‐based pruning (LAMP) for parameter reduction, achieving an optimal trade‐off between detection accuracy and computational efficiency under resource‐constrained conditions. Experimental results demonstrate superior performance on the underwater dataset. Compared to the YOLOv11 baseline, UGS‐YOLO_SP reduces parameters by 61.3% and computational costs (giga floating‐point operations per second, GFLOPs) by 39.7%, while achieving 98.66% precision and 98.18% mean average precision (mAP) at an intersection‐over‐union (IoU) threshold of 0.50 (mAP50).