EHL-YOLO: hierarchical multi-scale feature extraction and efficient cross-level fusion for UAV object detection

Small-object detection in unmanned aerial vehicle (UAV) imagery remains challenging because targets are small, densely distributed, and easily confused with complex backgrounds. This study presents EHL-YOLO, a task-oriented structural optimization of YOLOv10l that coordinates shallow feature preservation, hierarchical cross-scale interaction, and localization-quality-aware prediction. The backbone replaces selected C2f-based units with Efficient Adaptive Multi-Scale Feature Extraction (EAMFE), the neck reconstructs the P3–P5 fusion pathway through a Hierarchical Adaptive Feature Interaction (HAFI) design, and the detection head incorporates Location Quality Estimation and Dual-Branch Collaborative detection (LQEDC). On VisDrone-DET2019, EHL-YOLO increases Precision from 0.493 to 0.507, Recall from 0.374 to 0.402, mAP50 from 0.363 to 0.392, and mAP50–95 from 0.217 to 0.228, while reducing parameters from 24.317 M to 22.106 M and GFLOPs from 120.0 to 112.9. The expanded comparison includes RT-DETR-L and UAV-oriented detectors. Under its implementation-specific training and preprocessing configuration, RT-DETR-L obtains 0.3200 mAP50 and 0.1914 mAP50–95, whereas EHL-YOLO obtains 0.3920 and 0.2280, respectively. A Top-m sensitivity study shows that m = 4 provides the highest Recall, mAP50, and mAP50–95 among m∈{2,4,8}. Cross-dataset evaluation on the HIT-UAV infrared benchmark increases Precision from 0.8827 to 0.9221 but decreases Recall and mAP50–95, revealing a conservative prediction tendency and a remaining cross-modal limitation. These results demonstrate an improved accuracy–complexity balance for visible-light UAV detection under the reported protocol, while the HIT-UAV results reveal a remaining limitation under infrared domain shift.