EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Orientation-Aware Bilateral Feature Fusion Network for Small Object Detection in UAV Aerial Images

Zhixin Hu, Weimin Hou, Changlin Cui, Yanxia Wang

IEEE Access · 2026 · 14 · 112360-112376

DOI: 10.1109/access.2026.3715516 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Zhixin Hu und 3 weiteren Autor:innen. Die Arbeit nennt 61 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China, Major Science and Technology Support Project of Hebei Province, Natural Science Foundation of Hebei Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Small objects in unmanned aerial vehicle (UAV) aerial images usually occupy only a small proportion of pixels, exhibit weak local structures, and are densely distributed. They are also susceptible to complex backgrounds, occlusions, and scale variations. Meanwhile, changes in imaging viewpoints and differences in object motion directions make it difficult for fixed convolutional sampling grids to sufficiently match the edges, textures, and geometric structures of small and weak objects, further affecting feature extraction and localization quality. Existing methods mainly improve small object representation through shallow feature enhancement, multi-scale fusion, or attention mechanisms. However, several challenges remain, including insufficient modeling of orientation variations, inadequate interaction between deep and shallow features, and limited scale-differentiated decoding capability. To address these issues, this article proposes RBS-YOLO. First, a Cross-Stage Adaptive Rotated Convolution (CARC) module is designed to enhance local structural modeling for objects with orientation variations by adaptively adjusting the directions of convolution kernels. Second, a Bilateral Information Effective Fusion (BIEF) module is constructed to alleviate semantic inconsistency, background noise, and information redundancy in cross-layer fusion through bidirectional interaction between deep semantic information and shallow spatial details. Furthermore, a Size-Aware Decoder (SADecoder) is introduced to enhance detailed representation and semantic context modeling in detection branches according to the feature requirements of objects at different scales. On the VisDrone2019 dataset, RBS-YOLO improves Precision, Recall, mAP@0.5, and mAP@0.5:0.95 by 8.4, 5.6, 8.3, and 6.2 percentage points, respectively, compared with the baseline model. RBS-YOLO achieves 135.8 FPS with 3.1M parameters and 35.8 GFLOPs, maintaining real-time inference capability. In addition, experiments on the UAVDT, TinyPerson, and HRSC2016 datasets further verify the cross-scene adaptability of the proposed method.

Förderung

  • National Natural Science Foundation of China
  • Major Science and Technology Support Project of Hebei Province
  • Natural Science Foundation of Hebei Province
  • Hebei Province Higher Education Basic Scientific Research Operating Funds

Zitieren

APA

Hu, Z., Hou, W., Cui, C., & Wang, Y. (2026). Orientation-Aware Bilateral Feature Fusion Network for Small Object Detection in UAV Aerial Images. IEEE Access, 14, 112360-112376. https://doi.org/10.1109/access.2026.3715516

Vancouver

Hu Z, Hou W, Cui C, Wang Y. Orientation-Aware Bilateral Feature Fusion Network for Small Object Detection in UAV Aerial Images. IEEE Access. 2026;14:112360-112376. doi:10.1109/access.2026.3715516

BibTeX
@article{Hu2026,
  author = {Hu, Zhixin and Hou, Weimin and Cui, Changlin and Wang, Yanxia},
  title = {Orientation-Aware Bilateral Feature Fusion Network for Small Object Detection in UAV Aerial Images},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {112360--112376},
  doi = {10.1109/access.2026.3715516}
}
Kostenlos starten →