EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

SOFNet: A Structured Object Fusion Network for Object Detection in Complex Maritime Scenarios

Kehui Xu, Jiale Wang, Jianhua Cheng, Hongliang Yin, Jing Ma, Xihang Su, Meiting Liu

IEEE Access · 2026 · 14 · 100600-100616

DOI: 10.1109/access.2026.3706574 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Kehui Xu und 6 weiteren Autor:innen. Die Arbeit nennt 26 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Youth Talent Lifting Project of the China Association for Science and Technology.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

To address the severe challenges facing ship detection in complex maritime scenarios, such as complex background clutter interference, dense target occlusion, and large-scale variations, this paper proposes a lightweight yet high-precision Structured Object Fusion Network (SOFNet). Based on the YOLO11 architecture, the network constructs an optimization system ranging from low-level feature extraction to high-level semantic fusion, effectively enhancing the model’s robustness and discriminative capability under complex sea states. Three key modules are designed in this study. First, the Star-FE module, a lightweight feature extractor that employs a high-dimensional nonlinear mapping mechanism to bolster the representation of ship geometric topologies while minimizing computational overhead. Second, the C2PSA-EDFFN module integrates a frequency-domain feature recalibration mechanism to help reduce the influence of periodic maritime interference, such as waves and specular reflections, while enhancing target feature discriminability. Third, the CGA-Fusion module mitigates semantic misalignment in cross-scale fusion through adaptive attention weighting, thereby improving performance in densely clustered scenarios. Experimental results on the HRSC2016 benchmark dataset demonstrate that SOFNet achieves an mAP50 of 91.62% with only 2.9M parameters. On the SeaShips dataset, the model also delivers strong performance, attaining an overall mAP50–95 of 81.3%, demonstrating its robustness in multi-category and complex harbor environments. Furthermore, this study constructs a multi-view port simulation dataset, UE4-Port, based on Unreal Engine 4, and systematically evaluates model robustness from two dimensions: viewpoint variation and strong specular reflection interference.

Förderung

  • Youth Talent Lifting Project of the China Association for Science and Technology

Zitieren

APA

Xu, K., Wang, J., Cheng, J., Yin, H., Ma, J., Su, X., & Liu, M. (2026). SOFNet: A Structured Object Fusion Network for Object Detection in Complex Maritime Scenarios. IEEE Access, 14, 100600-100616. https://doi.org/10.1109/access.2026.3706574

Vancouver

Xu K, Wang J, Cheng J, Yin H, Ma J, Su X, et al. SOFNet: A Structured Object Fusion Network for Object Detection in Complex Maritime Scenarios. IEEE Access. 2026;14:100600-100616. doi:10.1109/access.2026.3706574

BibTeX
@article{Xu2026,
  author = {Xu, Kehui and Wang, Jiale and Cheng, Jianhua and Yin, Hongliang and Ma, Jing and Su, Xihang and Liu, Meiting},
  title = {SOFNet: A Structured Object Fusion Network for Object Detection in Complex Maritime Scenarios},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {100600--100616},
  doi = {10.1109/access.2026.3706574}
}
Kostenlos starten →