Related Experiment Videos
MSA-DET: A Multi-Scale Attention Network with Adaptive Feature Fusion for SAR Ship Detection.
Sai Wan1, Zhiyong Tao2, Lu Chen1
1School of Electronic and Information Engineering, Liaoning Technical University, Huludao 125105, China.
Sensors (Basel, Switzerland)
|July 15, 2026
Summary
This study introduces MSA-DET, an enhanced Synthetic Aperture Radar (SAR) ship detection network. MSA-DET improves detection accuracy for ships of all sizes by addressing speckle noise and background clutter.
Area of Science:
- Remote Sensing
- Computer Vision
- Artificial Intelligence
Background:
- Synthetic Aperture Radar (SAR) ship detection is challenged by speckle noise, complex backgrounds, and varying target sizes.
- Existing methods struggle to jointly address these issues effectively.
Purpose of the Study:
- To propose MSA-DET, an improved SAR ship detection network based on YOLOv11.
- To enhance the detection of ships with diverse spatial extents and orientations in SAR imagery.
Main Methods:
- Developed a Multi-Scale Cross-axis Attention (MSCAttention) module in the backbone for sharpened feature representation.
- Redesigned the C3k2 block to C3k2_SSA with sparse self-attention to suppress speckle and reduce computation.
- Implemented an Adaptive Spatial Feature Fusion (ASFF) detection head for improved scale-aware feature aggregation.
Main Results:
- MSA-DET achieved 63.6% mAP@0.5:0.95 and 88.1% mAP@0.5 on the HRSID dataset, outperforming YOLOv11n.
- On the SSDD dataset, MSA-DET reached 69.6% mAP@0.5:0.95 and 97.7% mAP@0.5, surpassing the baseline.
- Demonstrated significant accuracy gains with a moderate increase in model size and computational cost.
Conclusions:
- Coordinated multi-stage redesign is an effective strategy for SAR ship detection.
- MSA-DET offers a substantial improvement in detection quality, justifying the trade-off in computational resources.