RSM-YOLO: an object detection algorithm for marine debris

As the "Maritime Silk Road" initiative and the Regional Comprehensive Economic Partnership Agreement (RCEP) advance, global marine trade grows, resulting in a 12.7% yearly rise in marine plastic pollution and discarded fishing gear waste (UNEP, 2022). Currently, target detection of marine debris faces challenges such as complex underwater environments, significant scale variations of targets, and insufficient adaptability to deployment conditions. Given these limitations, this paper proposes an RSM-YOLO (You Only Look Once) model. Based on the YOLOv11 baseline model, RSM-YOLO incorporates a RepGhost-based backbone network (RepGhostNet), leveraging its lightweight and hardware-friendly characteristics to enhance deployment adaptability. Additionally, an improved SimAM attention mechanism is employed to better handle multi-scale large object detection in complex underwater scenarios. The model further optimizes the MPDIoU loss function for underwater multi-scale target detection scenarios and enhances detection sensitivity through an adaptive weighting mechanism, effectively addressing the challenge of significant target scale variations in underwater environments. Experimental findings show that RSM-YOLO outperforms various popular object detection algorithms, such as YOLOv5, YOLOv6, YOLOv8, YOLOv9, and YOLOv10, in terms of detection accuracy and overall performance. RSM-YOLO achieves a mAP@0.5 of 93.73% and an F1-score of 91%, demonstrating consistent excellence in detecting targets across different scales in challenging underwater environments. These outcomes highlight the exceptional precision of RSM-YOLO, offering a dependable method for identifying marine debris targets.