EUVIMEDEUVIMED entdecken
Crossref

LSMNet: A Lightweight Local-State Modeling Network for Tiny Object Detection in Aerial Images

Guangshuai Gao, Wei Zhang, Zhilin Zhang, Yan Dong, Jiangtao Xi

IEEE Geoscience and Remote Sensing Letters · 2026 · 23 · 6013205-6013205

DOI: 10.1109/lgrs.2026.3704444 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Geoscience and Remote Sensing Letters, 2026. Verfasst von Guangshuai Gao und 4 weiteren Autor:innen. Die Arbeit nennt 24 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Henan Key Research and Development Project, Youth Science Foundation Project (Category B) of the Natural Science Foundation of Henan Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Tiny object detection in aerial images has made significant progress; however, it remains constrained by two primary issues: 1) complex network architectures introduced to enhance feature representation capability improve accuracy but lead to a sharp increase in parameters and computational costs, making deployment on resource-constrained edge devices challenging and 2) existing lightweight designs, however, often struggle to capture fine-grained structures and long-range contextual dependencies required for tiny objects due to insufficient feature extraction capabilities, resulting in significantly degraded detection performance. The recently proposed Rep-Mamba model demonstrates efficient long-range modeling capabilities and low computation overhead in image super-resolution tasks through combining state-space modeling with reparameterized convolutions. However, when transferring it to the field of object detection, it may lead to issues such as ambiguous feature hierarchy and weakened spatial details. To address the aforementioned challenges, we propose LSMNet, a lightweight local-state modeling network. This architecture achieves task-oriented structural reconstruction of Rep-Mamba through a dedicated lightweight backbone network, LSM-TD, specifically designed for tiny object detection. It introduces the local-state modeling block (LSMBlock) module, which leverages a lightweight local state-space aggregation module (LSAM) to capture fine-grained local structures while modeling long-range spatial dependencies via an efficient state propagation paradigm. Experiments on the challenging AI-TOD-v2 and LEVIR-Ship dataset demonstrate that the proposed method achieves significantly reduced parameters and computational complexity with comparable detection accuracy. Code will be available at https://github.com/chibaobao000/lsmnet upon publication.

Förderung

  • National Natural Science Foundation of China
  • Henan Key Research and Development Project
  • Youth Science Foundation Project (Category B) of the Natural Science Foundation of Henan Province
  • Key Project of the Natural Science Foundation of Zhongyuan University of Technology

Zitieren

APA

Gao, G., Zhang, W., Zhang, Z., Dong, Y., & Xi, J. (2026). LSMNet: A Lightweight Local-State Modeling Network for Tiny Object Detection in Aerial Images. IEEE Geoscience and Remote Sensing Letters, 23, 6013205-6013205. https://doi.org/10.1109/lgrs.2026.3704444

Vancouver

Gao G, Zhang W, Zhang Z, Dong Y, Xi J. LSMNet: A Lightweight Local-State Modeling Network for Tiny Object Detection in Aerial Images. IEEE Geoscience and Remote Sensing Letters. 2026;23:6013205-6013205. doi:10.1109/lgrs.2026.3704444

BibTeX
@article{Gao2026,
  author = {Gao, Guangshuai and Zhang, Wei and Zhang, Zhilin and Dong, Yan and Xi, Jiangtao},
  title = {LSMNet: A Lightweight Local-State Modeling Network for Tiny Object Detection in Aerial Images},
  journal = {IEEE Geoscience and Remote Sensing Letters},
  year = {2026},
  volume = {23},
  pages = {6013205--6013205},
  doi = {10.1109/lgrs.2026.3704444}
}
Kostenlos starten →