EMSH-DETR: an efficient multi-scale and hybrid DETR for floating garbage detection

Abstract The increasing severity of global water pollution has made the detection of floating garbage on water surfaces crucial for ecological protection and water environment management. However, challenges such as the inconspicuous features of small objects, surface ripples, sunlight reflections, and background interference from shadows pose significant obstacles to existing object detection models. To address these issues, we propose EMSH-DETR, an efficient multi-scale hybrid Transformer-based object detection framework. First, we design a lightweight backbone network, hierarchical cooperative feature optimization, which enhances both local perception and global semantic modeling through the collaborative combination of the shallow-layer RAEF module and the deep-layer GLSGFA module. Second, an adaptive sparse attention mechanism, intra-scale feature interaction with adaptive sparse attention, is introduced to suppress ripple and reflection noise and enhance the model’s ability to focus on critical regions. Additionally, we construct a multi-branch fusion structure, EG-Fusion, to achieve more stable feature alignment and semantic integration in complex environments. Experiments conducted on the self-built Flow-L dataset demonstrate that EMSH-DETR achieves 90.8% mAP 50 and 62.2% mAP, outperforming RT-DETR by 2.6% and 1.6%, respectively. The model also reduces parameters and FLOPs by 30.2% and 23.7%, respectively, compared with the baseline. Furthermore, EMSH-DETR shows strong cross-domain generalization performance on the VisDrone2019 dataset. These results confirm that EMSH-DETR achieves higher accuracy and efficiency with lower computational cost, demonstrating strong practicality and deployment potential in intelligent water-surface garbage cleaning systems.