Efficient Underwater Garbage Detection Using GSConv Enhanced YOLOv8 with GD Mechanism

Abstract

Due to low visibility, scale variation, and complex backgrounds, detecting marine debris in underwater environments remains highly challenging. To address these issues, we propose YOLO-GGS, a lightweight yet high-performance object detector built upon YOLOv8. The framework incorporates three key innovations. First, the Gather-and-Distribute (GD) mechanism from Gold-YOLO is introduced into the neck, which unifies multi-scale feature aggregation while selectively injecting global context, thereby enhancing object perception across different scales. Second, GSConv-based hybrid convolutions are deployed in both the backbone and the injection module, effectively balancing rich channel interactions with reduced computational complexity. Third, a Slim-Neck design simplifies the feature fusion path by eliminating redundant operations, thus improving inference speed. Comprehensive experiments on the J-EDI and Brackish underwater
datasets demonstrate the superior performance of YOLO-GGS, achieving mAP@0.5:0.95 values of 88.5% and 84.7%, which represent improvements of 4% and 2.5% over the baseline model, respectively. Moreover, real-time evaluation shows that YOLO-GGS reaches an inference speed of 108.4 FPS. These results
highlight YOLO-GGS as an efficient and accurate solution for underwater debris detection, offering substantial potential for deployment on Autonomous Underwater Vehicles (AUVs) and Remotely Operated Vehicles (ROVs).