EUVIMEDEUVIMED entdecken
Crossref

SDR-YOLOv8: a recursive spatial-decoupled network for aerial tiny object detection

Lei Tian, Yanhao Guo, Furun Guo, Chengshen Lao, Zhanhao Yang

Physica Scripta · 03.07.2026 · 101 (26) · 266001

DOI: 10.1088/1402-4896/ae7edb ↗

Kurz gefasst

Fachartikel in der Zeitschrift Physica Scripta, online erschienen am 29.06.2026. Verfasst von Lei Tian und 4 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben. Gefördert von: Natural Science Basic Research Program of Shaanxi, Shaanxi Provincial Education Science Planning Project, of Xi.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The identification of very tiny objects in UAV aerial sensing photos is especially difficult, with the typical target dimension being just 12.8 pixels, far below that of conventional datasets. This leads to poor detection performance of existing models in scenarios with scale variations, background noise, and dense overlaps. In response to these issues, this study introduces the SDR-YOLOv8 framework, which integrates multiple targeted optimizations. It builds a recursive attention feature pyramid structure to accomplish gradual multi-level integration and residue adjustment, enhancing the retention of shallow-level details. It develops a spatio-temporal dynamic convolution module that combines a three-branch convolution and MLP density guidance to adaptively adjust the response to irregular shapes and density changes. It designs a dynamic channel modulation attention mechanism that extends the coordinate attention framework to strengthen the feature representation in high-density regions. It introduces a composite loss of normalized Wasserstein distance + Wise-IoU + distribution focus loss, combining scale matching and distribution focus to improve boundary regression accuracy and noise robustness. These innovations collectively enhance the model’s adaptability to complex aerial remote sensing environments while keeping the parameter count within 9.9 MB for efficient operation. On the AI-TOD dataset, SDR-YOLOv8 achieves an mAP@0.5 of 42.8%, a 4.5% improvement over YOLOv8s; on the VisDrone2019 dataset, it reaches an mAP@0.5 of 46.5%, a 7.2% increase. It performs stably across different IoU thresholds. Ablation experiments verify the synergistic effect of each component, and visualization results show a significant reduction in missed detections in both normal and occluded scenarios. This algorithm contributes a practical framework for real-time remote sensing applications such as border monitoring and disaster response.

Förderung

  • Natural Science Basic Research Program of Shaanxi
  • Shaanxi Provincial Education Science Planning Project
  • of Xi

Zitieren

APA

Tian, L., Guo, Y., Guo, F., Lao, C., & Yang, Z. (2026). SDR-YOLOv8: a recursive spatial-decoupled network for aerial tiny object detection. Physica Scripta, 101(26), 266001. https://doi.org/10.1088/1402-4896/ae7edb

Vancouver

Tian L, Guo Y, Guo F, Lao C, Yang Z. SDR-YOLOv8: a recursive spatial-decoupled network for aerial tiny object detection. Physica Scripta. 2026;101(26):266001. doi:10.1088/1402-4896/ae7edb

BibTeX
@article{Tian2026,
  author = {Tian, Lei and Guo, Yanhao and Guo, Furun and Lao, Chengshen and Yang, Zhanhao},
  title = {SDR-YOLOv8: a recursive spatial-decoupled network for aerial tiny object detection},
  journal = {Physica Scripta},
  year = {2026},
  volume = {101},
  number = {26},
  pages = {266001},
  doi = {10.1088/1402-4896/ae7edb}
}
Kostenlos starten →