Abstract
This paper presents a novel approach to Synthetic Aperture Radar (SAR) ship detection, addressing the limitations of static label assignment and the misalignment between classification confidence and localization accuracy. We introduce the Prediction and Prior Information Guided (PPIG) label assignment strategy and the Multiscale Ship Detection Network (MSDNet). PPIG adaptively selects training samples by integrating prior overlap with real-time model predictions, including a dedicated IoU prediction branch to refine Non-Maximum Suppression (NMS). MSDNet enhances small target detection through Multilevel Feature Fusion (MLFF) and Adaptively Spatial Feature Fusion (ASFF). Evaluated on SSDD and HRSID datasets, our method achieves state-of-the-art performance (94.05% AP on SSDD) with reduced computational complexity compared to standard models like YOLO v4.
Introduction
Synthetic Aperture Radar (SAR) sensors are indispensable for maritime surveillance due to their all-weather, day-and-night imaging capabilities. As the field has evolved from traditional statistical methods like Constant False Alarm Rate (CFAR) to modern deep learning architectures, new challenges have emerged regarding detection precision and efficiency.
Current deep learning models face three primary hurdles:
- Label Assignment Bottleneck: Traditional fixed-rule assignment (e.g., static IoU thresholds) fails to account for the model's learning progress.
- Training-Testing Mismatch: High confidence scores often do not correlate with high localization accuracy, leading to suboptimal results during Non-Maximum Suppression (NMS).
- Small Target Detection: Ships in SAR imagery are often only a few pixels wide, making them difficult to distinguish from sea clutter.
Methodology
Prediction and Prior Information Guided (PPIG) Label Assignment
To address the limitations of static rules, we propose the PPIG strategy. Instead of a fixed threshold, PPIG calculates a dynamic "quality score" for candidate anchors. This score combines prior spatial information (initial overlap) with the model's current learning status, specifically its predicted confidence and a newly introduced predicted IoU branch.
By treating IoU prediction as a primary task, the model gains "self-awareness" regarding its localization precision. This ensures that during the NMS phase, the most accurately positioned boxes are retained, rather than just the most "confident" ones.
Multiscale Ship Detection Network (MSDNet)
To improve the detection of small targets, MSDNet incorporates two key architectural enhancements:
- Multilevel Feature Fusion (MLFF): This module adds a finer-grained feature layer (P3) and utilizes feature reuse branches to preserve the detailed textures of small ships that are often lost in deeper network layers.
- Adaptively Spatial Feature Fusion (ASFF): ASFF learns spatial weights to fuse features across different scales, allowing the model to focus on the most discriminative clues for ships of varying sizes while ignoring contradictory information.
Experimental Results
The proposed method was evaluated on two benchmark datasets: the SAR Ship Detection Dataset (SSDD) and the High-Resolution SAR Images Dataset (HRSID).
- Accuracy: On the SSDD dataset, the combination of MSDNet and PPIG achieved an Average Precision (AP) of 94.05%, a significant improvement over the YOLO v4 baseline (86.05%).
- Small Ship Performance: On the more challenging HRSID dataset, the method reached 86.11% AP, outperforming YOLO v5l by over 10%.
- Efficiency: Despite its advanced fusion modules, MSDNet is computationally efficient, requiring 8.37% fewer FLOPs than YOLO v4.
Conclusion and Future Work
The PPIG strategy and MSDNet architecture provide a robust solution for SAR ship detection, particularly for small targets in complex environments. By dynamically aligning label assignment with model performance and enhancing multiscale feature preservation, we achieve state-of-the-art results with high efficiency. Future research will focus on improving discrimination in high-clutter coastal environments to reduce false positives from land structures and refining detection for targets at extreme scales.
Summary updated