EUVIMEDEUVIMED entdecken
Crossref

SAR-DETR: Multiscale Granularity Guided End-to-End SAR Target Detection Transformer

Yimin Zhang, Chuxuan Chen, Chuhao Ruan, Lingfei Ma, Haiyan Guan, Yongtao Yu, Jonathan Li

IEEE Transactions on Aerospace and Electronic Systems · 2026 · 62 · 13648-13663

DOI: 10.1109/taes.2026.3710962 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Aerospace and Electronic Systems, 2026. Verfasst von Yimin Zhang und 6 weiteren Autor:innen. Die Arbeit nennt 57 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Natural Science Foundation of Jiangsu Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Recently, deep learning-based synthetic aperture radar (SAR) target detection methods significantly improve the target recognition capability in complex scenes. However, due to speckle noise, geometric distortions, and scale variations of targets, existing SAR target detectors cannot well balance detection accuracy and computational complexity. Most existing Transformer-based SAR detectors simply adapt optical detection structures without explicitly utilizing SAR target geometry prior, leading to limited performance in complex scenes. To address these challenges, we propose SAR-DETR, a multiscale granularity guided end-to-end Transformer designed specifically for SAR target detection. First, MG$^{2}$A is proposed to explicitly embed SAR target geometry prior into visual attention, combining dynamic threshold generation and spatial attention to enhance feature discriminability. Then, a backbone high-level feature fusion and large kernel network (HF$^{2}$LKNet) is introduced to expand receptive fields and suppress geometric distortions, using star operations and large-kernel convolutions. Finally, the adaptive local feature compensation (ALFC) module is proposed to compensate local details weakened by self-attention, further improving detection robustness. Extensive experiments on SARDet-100 k and HRSID show that SAR-DETR achieves state-of-the-art performance with mAP 0.586 and 0.728, respectively, while maintaining a compact architecture of 22.94 M parameters.

Förderung

  • National Natural Science Foundation of China
  • Natural Science Foundation of Jiangsu Province

Zitieren

APA

Zhang, Y., Chen, C., Ruan, C., Ma, L., Guan, H., Yu, Y., & Li, J. (2026). SAR-DETR: Multiscale Granularity Guided End-to-End SAR Target Detection Transformer. IEEE Transactions on Aerospace and Electronic Systems, 62, 13648-13663. https://doi.org/10.1109/taes.2026.3710962

Vancouver

Zhang Y, Chen C, Ruan C, Ma L, Guan H, Yu Y, et al. SAR-DETR: Multiscale Granularity Guided End-to-End SAR Target Detection Transformer. IEEE Transactions on Aerospace and Electronic Systems. 2026;62:13648-13663. doi:10.1109/taes.2026.3710962

BibTeX
@article{Zhang2026,
  author = {Zhang, Yimin and Chen, Chuxuan and Ruan, Chuhao and Ma, Lingfei and Guan, Haiyan and Yu, Yongtao and Li, Jonathan},
  title = {SAR-DETR: Multiscale Granularity Guided End-to-End SAR Target Detection Transformer},
  journal = {IEEE Transactions on Aerospace and Electronic Systems},
  year = {2026},
  volume = {62},
  pages = {13648--13663},
  doi = {10.1109/taes.2026.3710962}
}
Kostenlos starten →