EUVIMEDEUVIMED entdecken
Crossref

CMDNet: A Compact Detail-Preserving Network for UAV Vision-Based Small Object Detection

Zhihui Hu, Chiyao Mo, Yanxiong Li

IEEE Transactions on Instrumentation and Measurement · 2026 · 75 · 5020012-5020012

DOI: 10.1109/tim.2026.3742029 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Instrumentation and Measurement, 2026. Verfasst von Zhihui Hu und 2 weiteren Autor:innen. Die Arbeit nennt 58 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Exchange Project of the 10th Meeting of China-Croatia Science and Technology Cooperation Committee.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

With the continuous advancement of unmanned aerial vehicle (UAV) technology, vision-based object detection has become an important component of UAV sensing, monitoring, and measurement-oriented perception systems. Due to complex backgrounds, dense distributions of small objects, and the need for lightweight model design, accurate and efficient detection of small objects in UAV images remains challenging. To address these challenges, in CMDNet, we redesign the backbone and detection heads by reducing the number of backbone downsampling stages and removing the detection head responsible for large objects. This design preserves more fine-grained information for small objects while reducing the parameter count and encourages the model to focus on small and medium objects. The four-stage backbone is adopted as an accuracy-efficiency tradeoff rather than an accuracy-optimal configuration. In addition, the convolution shuffle space to depth (CSSD) and multiscale shuffle space to depth (MSSD) modules are designed to enlarge the model’s receptive field and retain more detailed features of tiny objects during downsampling with limited computational overhead. We further develop a lightweight multibranch fusion pyramid network (MBFPN) as the feature-fusion module of CMDNet. It efficiently fuses deep semantic and shallow spatial features to improve small-object detection accuracy. Moreover, the scale- and task-aware head (ST-Head) is incorporated to strengthen scale- and task-aware prediction, thereby enhancing detection precision in UAV imagery. Experimental results show that CMDNet achieves AP0.5 scores of 48.2 %, 46.0 %, and 76.2 % on the VisDrone2021, AI-TODv2, and SeaDronesSeeV2 datasets, respectively. It outperforms the corresponding baselines by 10.3 %, 12.6 %, and 15.2 %, while maintaining low model complexity with 3.02M parameters and 48.1 GFLOPs.

Förderung

  • National Natural Science Foundation of China
  • Exchange Project of the 10th Meeting of China-Croatia Science and Technology Cooperation Committee

Zitieren

APA

Hu, Z., Mo, C., & Li, Y. (2026). CMDNet: A Compact Detail-Preserving Network for UAV Vision-Based Small Object Detection. IEEE Transactions on Instrumentation and Measurement, 75, 5020012-5020012. https://doi.org/10.1109/tim.2026.3742029

Vancouver

Hu Z, Mo C, Li Y. CMDNet: A Compact Detail-Preserving Network for UAV Vision-Based Small Object Detection. IEEE Transactions on Instrumentation and Measurement. 2026;75:5020012-5020012. doi:10.1109/tim.2026.3742029

BibTeX
@article{Hu2026,
  author = {Hu, Zhihui and Mo, Chiyao and Li, Yanxiong},
  title = {CMDNet: A Compact Detail-Preserving Network for UAV Vision-Based Small Object Detection},
  journal = {IEEE Transactions on Instrumentation and Measurement},
  year = {2026},
  volume = {75},
  pages = {5020012--5020012},
  doi = {10.1109/tim.2026.3742029}
}
Kostenlos starten →