A scale‑aware feature enhancement and localization refinement network for UAV small object detection

Abstract UAV-based small object detection remains challenging because aerial targets are often densely distributed, weakly textured, partially occluded, and affected by complex backgrounds and illumination variations. To address these issues, this paper presents SAFLR-Net, a scale-aware feature enhancement and localization refinement network for UAV small object detection. Rather than relying on simple model scaling, SAFLR-Net improves small-object perception through coordinated feature representation, high-resolution fusion, adaptive prediction, and localization refinement. Specifically, MSCM-C3k2 enhances local multi-scale contextual representation for weak-texture targets through lightweight multi-branch convolution and channel mixing. HSAF-Neck introduces a high-resolution P2 branch, DySample-based upsampling, and learnable weighted fusion to preserve fine spatial cues and strengthen cross-scale interaction. A DyHead-based prediction head further improves feature adaptability across scale, spatial, and task dimensions, while IFIoU combines auxiliary-box supervision with adaptive IoU remapping to improve localization stability for small objects with ambiguous boundaries. Experiments on multiple aerial small-object detection datasets demonstrate the effectiveness of the proposed framework. On VisDrone2019, SAFLR-Net achieves 46.1% mAP50 and 28.3% mAP50:95, outperforming YOLO11n by 13.0 and 8.8 percentage points, respectively. On DroneVehicle, it obtains 72.7% mAP50 and 51.9% mAP50:95. Additional dataset-specific experiments on SODA-A and AI-TOD further show that the proposed architectural improvements remain effective across aerial datasets with different object-scale distributions and scene characteristics. These results demonstrate that SAFLR-Net provides stronger small-object representation and localization while maintaining a compact parameter scale for challenging UAV detection scenarios.