A Lightweight Multi-Scale Attention-Based Architecture Based on YOLO11 for Ship Detection in Synthetic Aperture Radar Images
A Lightweight Multi-Scale Attention-Based Architecture Based on YOLO11 for Ship Detection in Synthetic Aperture Radar Images
Shaghayegh Hooshyar1 Sadegh Samadi2 Shahrzad Hooshyar3 Matin Kamani4 Mohsen Eslami5
1) M.Sc. Student in Electrical Engineering, School of Electrical Engineering, Shiraz University of Technology, Shiraz, Iran —
2) Associate Professor, School of Electrical Engineering, Shiraz University of Technology, Shiraz, Iran —
3) M.Sc. Student in Electrical Engineering, School of Electrical Engineering, Shiraz University of Technology, Shiraz, Iran —
4) Ph.D. Student in Electrical Engineering, School of Electrical Engineering, Shiraz University of Technology, Shiraz, Iran —
5) Associate Professor, School of Electrical Engineering, Shiraz University of Technology, Shiraz, Iran —
IOI Code
:
AAHD-ABHEBB
Publication :
National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
Ship detection in Synthetic Aperture Radar (SAR) imagery remains a significant challenge in maritime surveillance systems due to the presence of speckle noise, large variations in target scale, the prevalence of small vessels, complex maritime scenes, and the limitations of conventional network architectures in effective multi-scale feature extraction and fusion. In this paper, several architectural modifications are introduced to the YOLO11s model to enhance its ship detection performance in SAR imagery. These modifications are designed to strengthen feature extraction and multi-scale feature fusion, improve the detection capability for small and multi-scale targets, and reduce computational complexity. While preserving the computational efficiency of the baseline model, the proposed architecture enhances both detection accuracy and target localization quality, achieving a more favorable trade-off between detection performance and computational cost. Experimental evaluation on the HRSID dataset demonstrates that the proposed method improves the mAP from 89.71% for the baseline model to 90.74%. Moreover, the number of model parameters is reduced by 19.1% compared with the baseline network. The proposed architecture exhibits superior performance in detecting small and multi-scale ships as well as vessels located in complex maritime environments, while maintaining computational efficiency, making it a promising candidate for real-time maritime monitoring and surveillance applications.
Keywords : ship detection sar images deep learning small target detection multi scale feature fusion yolo11s