AE-DETR: An Accuracy–Efficiency Balanced DETR for Water-Surface Floating Garbage Detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Image-based two-dimensional visual measurement enables non-contact acquisition of target categories and image-plane locations for automated aquatic environmental monitoring. Water-surface floating-garbage detection remains challenging because water ripples, reflections, and illumination variations create complex backgrounds, while distant small targets occupy few pixels and provide limited feature information, reducing recognition and image-plane localization reliability. Deployment on mobile edge platforms such as unmanned surface vehicles also constrains inference speed, storage overhead, and computational cost. To address these issues, this paper proposes the Accuracy–Efficiency Balanced DETR (AE-DETR), based on RT-DETR, to balance detection accuracy and efficiency. Progressive Guided Sparse Attention (PGSA) is introduced as the principal methodological contribution and efficiency-oriented core, reducing redundant encoder interactions and normalization overhead while preserving effective global contextual modeling. A robustness-oriented lightweight backbone, the Feature-Adaptive Dynamic Aggregation Network (FADA-Net), improves target-feature representation under directional water-surface interference and illumination variations. An accuracy-oriented Small-Object Optimized Spatial–Frequency Fusion Pyramid (SOSF-FP) enhances the overall representation of small floating objects in complex scenes. On the self-constructed Flow-D dataset, AE-DETR improves mAP50 from 85.2% to 89.0% and mAP50–95 from 58.3% to 63.9%, while reducing parameters from 19.9 M to 14.0 M. This gain is accompanied by a modest FLOPs increase from 57 G to 59 G and a slight speed decrease from 174.0 to 165.0 FPS. Supplementary experiments on VisDrone2019 examine performance under a different UAV-view small-object distribution without serving as direct evidence of generalization to diverse water-surface conditions. TensorRT FP16 mixed-precision deployment on an NVIDIA Jetson Orin Nano achieves 88.2% mAP50, 63.4% mAP50–95, an engine-only latency of 14.2 ms, and 70.5 FPS. Qualitative on-water validation further confirms operation of the onboard detection pipeline in real-world scenes. Overall, AE-DETR provides a practical accuracy–efficiency balanced solution for resource-constrained mobile platforms and related outdoor mobile visual-monitoring tasks.