Abstract
There are several unresolved issues in the field of ship instance segmentation in synthetic aperture radar (SAR) images. First, in inshore dense ship area, the problems of missed detections and mask overlap frequently occur. Second, in inshore scenes, false alarms occur due to strong clutter interference. In order to address these issues, we propose a novel ship instance segmentation network based on dynamic key points information enhancement. In the detection branch of the network, a dynamic key points module is designed to incorporate the target's geometric information into the parameters of the dynamic mask head using an implicit encoding technique. In addition, we introduce a dynamic key points encoding branch, which encodes the target's strong scattering regions as dynamic key points. It strengthens the network's ability to learn the correspondence between local regions with strong scattering and overall ship targets, effectively mitigating mask overlap issues. Moreover, it enhances the discriminative ability of network between ship targets and clutter interference, leading to a reduction in false alarm rates. To further enhance the dynamic key points information, an instancewise attention map module is designed, which decodes the key points during the mask prediction period, generating instancewise attention maps based on 2-D Gaussian distribution. This module further enhances the sensibility of network to specific instances. Simulation experiments conducted on the Polygon Segmentation SAR Ship Detection Dataset and High-Resolution SAR Images Dataset demonstrate the superiority of our proposed method over other state-of-the-art methods in inshore and offshore scenes.
1. Introduction
Synthetic Aperture Radar (SAR) is a critical tool for maritime surveillance due to its all-weather and day-and-night imaging capabilities. However, ship instance segmentation in SAR images remains a complex task. Unlike optical imagery, SAR is prone to speckle noise and strong clutter from land-based structures like docks and ports. Current segmentation methods struggle with two primary challenges: inshore confusion, where tightly packed ships lead to merged masks or missed detections, and false alarms, where land-based metal structures are incorrectly identified as vessels.
2. Methodology
The proposed network is built upon the CondInst framework, enhanced with two novel information-processing modules designed to leverage the physical scattering properties of ships.
2.1 Dynamic Key Points Module (DKPM)
The DKPM identifies key points based on the ship's geometry, specifically focusing on the ship's bow—a region that typically produces strong, distinct radar returns. Using a symmetric non-uniform sampling strategy, these geometric features are implicitly encoded directly into the parameters of the mask prediction head, allowing the network to adapt its vision dynamically for every instance.
2.2 Instance-wise Attention Map Module (IAMM)
This module acts as a spatial filter by decoding key point information into a 2D Gaussian distribution heatmap for every detected instance. This heatmap serves as a spatial attention mask, directing the network to focus on a specific ship's pixels while suppressing interference from nearby ships or land clutter.
2.3 Angle Encoding
To handle ships at various orientations, the authors utilize Dense Label Encoding (DCL). By treating angle prediction as a classification problem rather than regression, the model avoids the mathematical discontinuities that occur during ship rotation.
3. Results and Discussion
The model was evaluated on two major benchmarks: the PSeg-SSDD (Polygon Segmentation SAR Ship Detection Dataset) and the HRSID (High-Resolution SAR Images Dataset).
Experimental results demonstrate that the proposed method consistently outperforms state-of-the-art models like Mask R-CNN and SOLOv2. In inshore scenes on the PSeg-SSDD dataset, the model achieved an AP of 0.628. It showed particular strength in the AP75 metric, reaching 0.935 in offshore scenes, indicating highly accurate mask boundaries. The integration of DKPM and IAMM significantly reduced false alarms and improved the detection of small objects in crowded port environments.
4. Conclusion
By incorporating dynamic key points and instance-wise attention, this research successfully addresses the persistent issues of mask overlap and clutter-induced false alarms in SAR ship segmentation. While the model increases computational complexity—with approximately 40% more FLOPs than the baseline—the gain in segmentation fidelity in complex inshore environments provides a robust foundation for future real-time maritime surveillance applications.
Summary updated