TY  - JOUR
AU  - Wang, Wenqi
AU  - Zhang, Xu
AU  - Ma, Jun
AU  - Ren, Xunhuan
AU  - Tsviatkou, Viktar Yurevich
PY  - 2026
DA  - 2026/06/11
TI  - An Improved Yolov12-Based Object Detection Model For Ship Monitoring in SAR Images
JO  - Chinese Journal of Information Fusion
T2  - Chinese Journal of Information Fusion
JF  - Chinese Journal of Information Fusion
VL  - 3
IS  - 2
SP  - 125
EP  - 137
DO  - 10.62762/CJIF.2025.869982
UR  - https://www.icck.org/article/abs/CJIF.2025.869982
KW  - YOLOv12
KW  - ship detection
KW  - small-scale target
KW  - synthetic aperture radar (SAR)
AB  - Ship detection in Synthetic Aperture Radar (SAR) imagery is crucial for maritime surveillance. However, it faces significant challenges, including small target sizes, complex sea clutter interference, and stringent requirements for computational efficiency in on-board processing. While detection frameworks like YOLOv12 have achieved a favorable balance between speed and accuracy by integrating attention mechanisms with convolutional neural networks (CNNs), their generic architectures are not optimized for the unique physical characteristics of SAR imagery and the scattering properties of ship targets. To develop a more suitable lightweight and high-precision model for SAR ship detection, this study proposes an improved YOLOv12 framework. Specifically, two modules are adopted: First, the GhostStem module is embedded into the shallow network layers to replace traditional convolutional layers. This lightweight feature extraction module effectively reduces the number of parameters and computational cost in the early stages, establishing an efficient foundation for target detection in SAR images. Second, the OverLookGate (OLGate) module is incorporated. By extracting lightweight global semantic priors and employing a two-level feature gating mechanism, it significantly enhances the model's capability to discriminate and localize features within SAR imagery under complex backgrounds (e.g., coastlines and island interference) and among distributed small-scale ship targets. Experiments on publicly available SAR ship detection datasets show that, compared with the original YOLOv12 and other mainstream detectors, the proposed improved model maintains high accuracy while demonstrating competitive performance, particularly achieving significant improvements in Recall and [email protected], especially in achieving higher recall and overall accuracy for small targets in complex scenarios.
SN  - 2998-3371
PB  - Institute of Central Computation and Knowledge
LA  - English
ER  - 

@article{Wang2026An,
  author = {Wenqi Wang and Xu Zhang and Jun Ma and Xunhuan Ren and Viktar Yurevich Tsviatkou},
  title = {An Improved Yolov12-Based Object Detection Model For Ship Monitoring in SAR Images},
  journal = {Chinese Journal of Information Fusion},
  year = {2026},
  volume = {3},
  number = {2},
  pages = {125-137},
  doi = {10.62762/CJIF.2025.869982},
  url = {https://www.icck.org/article/abs/CJIF.2025.869982},
  abstract = {Ship detection in Synthetic Aperture Radar (SAR) imagery is crucial for maritime surveillance. However, it faces significant challenges, including small target sizes, complex sea clutter interference, and stringent requirements for computational efficiency in on-board processing. While detection frameworks like YOLOv12 have achieved a favorable balance between speed and accuracy by integrating attention mechanisms with convolutional neural networks (CNNs), their generic architectures are not optimized for the unique physical characteristics of SAR imagery and the scattering properties of ship targets. To develop a more suitable lightweight and high-precision model for SAR ship detection, this study proposes an improved YOLOv12 framework. Specifically, two modules are adopted: First, the GhostStem module is embedded into the shallow network layers to replace traditional convolutional layers. This lightweight feature extraction module effectively reduces the number of parameters and computational cost in the early stages, establishing an efficient foundation for target detection in SAR images. Second, the OverLookGate (OLGate) module is incorporated. By extracting lightweight global semantic priors and employing a two-level feature gating mechanism, it significantly enhances the model's capability to discriminate and localize features within SAR imagery under complex backgrounds (e.g., coastlines and island interference) and among distributed small-scale ship targets. Experiments on publicly available SAR ship detection datasets show that, compared with the original YOLOv12 and other mainstream detectors, the proposed improved model maintains high accuracy while demonstrating competitive performance, particularly achieving significant improvements in Recall and [email protected], especially in achieving higher recall and overall accuracy for small targets in complex scenarios.},
  keywords = {YOLOv12, ship detection, small-scale target, synthetic aperture radar (SAR)},
  issn = {2998-3371},
  publisher = {Institute of Central Computation and Knowledge}
}