Improved Transformer-based detection of underwater plastic debris in complex environments

IntroductionUnderwater plastic debris detection remains challenging in visually cluttered environments because underwater images are often severely degraded, debris instances are small, and backgrounds are complex. To address these challenges, this study develops an improved RF-DETR detector for underwater plastic debris detection.MethodsThe proposed architecture integrates an underwater frequency-aware feature reweighting module, a multi-scale frequency-aware attention module, and a query adaptive reweighting module with bounded learnable scales. These modules are designed to enhance the representation of weak-texture and low-contrast features while improving query-level localization and classification. The model was evaluated on the TrashCan dataset and further assessed through a transfer experiment on the DeepTrash dataset.ResultsOn the TrashCan dataset, the proposed detector achieved 93.11% precision, an F1-score of 86.63%, an mAP@50 of 91.84%, and an mAP@50:95 of 69.92%. In the DeepTrash transfer experiment, the TrashCan-trained checkpoint was adapted to the one-class plastic target dataset and achieved 91.67% precision, 79.00% recall, an F1-score of 84.86%, an mAP@50 of 86.23%, and an mAP@50:95 of 55.06%. Compared with representative baselines, including YOLOv11n, YOLO26n, YOLO26s, and RT-DETR, the improved RF-DETR achieved the strongest overall performance in the reported TrashCan evaluation and DeepTrash transfer experiment. Ablation results further showed that the three proposed modules performed best when used in combination.DiscussionThe results demonstrate that the proposed frequency-aware feature enhancement and adaptive query reweighting strategies improve the detection of small and low-contrast underwater debris. The transfer results also indicate promising cross-dataset generalization. Overall, the proposed method provides a favorable trade-off between detection performance and model complexity on the evaluated datasets.