SL-YOLO: Scale- and Lighting-Aware Object Detection for Challenging UAV Imagery

Abstract

Complex illumination, perspective-induced scale variation, and repeated downsampling degrade object representations in unmanned aerial vehicle (UAV) imagery and make small-object localization highly sensitive to pixel-level shifts. To address these issues, we propose Scale- and Lighting-Aware YOLO (SL-YOLO), a lightweight object detector built on YOLOv11n. SL-YOLO integrates three complementary mechanisms. First, illumination-aware modulation compensates for feature-response degradation caused by low illumination, shadows, and local overexposure. Second, cross-branch scale interaction models dependencies among receptive-field branches to coordinate local details and global semantics. Third, a dedicated high-resolution small-object detection pathway combines shallow-detail enhancement, dynamic upsampling, a P2 high-resolution detection branch, and a distribution-distance loss. Together, these components link feature extraction, cross-level alignment, fine-grained prediction, and bounding-box regression. On the VisDrone2019-DET test-dev set, SL-YOLO achieves 33.5% $\mathrm {mAP}_{50}$ and 18.6% $\mathrm {mAP}_{50:95}$ with 2.94M parameters and 11.8 GFLOPs. These results exceed those of YOLOv11n by 5.8 and 3.2 percentage points, respectively. Progressive ablation and module-replacement experiments demonstrate the effectiveness of the proposed mechanisms, while class-wise and scale-specific analyses show consistent improvements for small objects. When trained independently under the same protocol, SL-YOLO also outperforms YOLOv11n on AI-TOD and DIOR-HBB. Overall, with fewer than 3M parameters, SL-YOLO achieves detection accuracy close to that of larger detector variants at the s scale and demonstrates a favorable accuracy–complexity balance in challenging UAV scenes.