CER-DETR: Channel-aware and edge-aligned detection of floating debris under complex weather

Accurate detection and removal of floating debris on water surfaces are crucial for ecological conservation and water management; however, existing methods suffer from limited generalization capabilities under unpredictable weather conditions. To address this, this study proposes a lightweight single-stage detection framework, Channel-aware Lightweight Edge-Align Residual-gated DETR (CER-DETR), designed to adaptively handle imaging degradation caused by complex meteorological conditions. The framework proposes a Lightweight Channel-Aware Residual Block (LCABlock) in the backbone network during downsampling to achieve channel selectivity and downsampling fidelity, thereby enhancing the signal-to-noise ratio for small objects. In global modeling, the Gated Residual Attention Block (GRAB) is designed to incorporate channel gating and dynamic residual relabeling, suppressing surface texture and background leakage. Finally, Scale Align EdgeFusion (SAEF) implements adaptive weighted fusion with cross-scale alignment. This study conducts systematic experiments using Flow-Img, River float, and a synthetic complex weather subset comprising Flow-lowlight, Flow-rain, Flow-foggy, Flow-mixed and HMHTT-mixed. Our method achieves mAP50 = 87.23 % on the Flow dataset with only 11.8 M. On four complex weather subsets, the mAP50 values are 85.13 %, 87.53 %, 85.39 %, and 85.99 %, respectively. Experimental and visualization analyses demonstrate that this approach exhibits significant advantages in the accuracy-efficiency tradeoff, outperforming multiple mainstream comparison methods overall. It can provide an important contribution to water environment monitoring and protection.