EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

VN-YOLO: Detail-Preserving and Recalibrated Multi-Scale YOLO for Small Objects in UAV Imagery

Ngoc-Tuan Hoang, Cao-Phuc Ha, Phu-Nguyen Le

IEEE Access · 2026 · 14 · 87951-87960

DOI: 10.1109/access.2026.3697316 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Ngoc-Tuan Hoang und 2 weiteren Autor:innen. Die Arbeit nennt 42 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Uncrewed aerial vehicles (UAVs) have fundamentally altered many fields due to their flexibility and ability to reach inaccessible areas. They are used in military reconnaissance, agricultural operations, disaster response, environmental monitoring, and traffic surveillance. Nevertheless, identifying diminutive objects in UAV imagery remains a persistent challenge due to low resolution, complex backgrounds, and hardware limitations. Conventional object detection algorithms, including two-stage approaches such as Faster R-CNN and one-stage detectors such as YOLO, often face challenges in these scenarios, particularly when real-time performance is essential. Consequently, this study presents VN-YOLO, a modified version of YOLOv11n, specifically designed to detect small objects in drone imagery. The contributions include: 1) a detail-preserving backbone that incorporates SPDConv (Space-to-Depth Convolution), Learnable Adaptive Downsampling (LAD), RepVGG, SwinTransformer, and SPPF_LSKA for enhanced feature extraction; 2) a robust neck with C2fFast, CBAM/RFA attention, ReCalibrationFPN, AMFE, and DCNv3 for multi-scale fusion and deformation handling; and 3) a stable detection head with channel alignment and MPDIoU loss for improved localization. Empirical findings derived from the VisDrone dataset demonstrate that VN-YOLO attains a mAP50 of 0.52 and a mAP50:95 of 0.356, surpassing the baseline YOLOv11n (mAP50 of 0.3314) by 57%, all the while preserving computational efficacy. Thus, this study advances real-time small-object detection in UAV applications by introducing a model that balances accuracy and efficiency.

Zitieren

APA

Hoang, N., Ha, C., & Le, P. (2026). VN-YOLO: Detail-Preserving and Recalibrated Multi-Scale YOLO for Small Objects in UAV Imagery. IEEE Access, 14, 87951-87960. https://doi.org/10.1109/access.2026.3697316

Vancouver

Hoang N, Ha C, Le P. VN-YOLO: Detail-Preserving and Recalibrated Multi-Scale YOLO for Small Objects in UAV Imagery. IEEE Access. 2026;14:87951-87960. doi:10.1109/access.2026.3697316

BibTeX
@article{Hoang2026,
  author = {Hoang, Ngoc-Tuan and Ha, Cao-Phuc and Le, Phu-Nguyen},
  title = {VN-YOLO: Detail-Preserving and Recalibrated Multi-Scale YOLO for Small Objects in UAV Imagery},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {87951--87960},
  doi = {10.1109/access.2026.3697316}
}
Kostenlos starten →