Abstract
Common horizontal bounding box (HBB)-based methods are not capable of accurately locating slender ship targets with arbitrary orientations in synthetic aperture radar (SAR) images. Therefore, in recent years, methods based on oriented bounding box (OBB) have gradually received attention from researchers. However, most of the recently proposed deep learning-based methods for OBB detection encounter the boundary discontinuity problem in angle or key point regression. In order to alleviate this problem, researchers propose to introduce some manually set parameters or extra network branches for distinguishing the boundary cases, which make training more difficult and lead to performance degradation. In this paper, in order to solve the boundary discontinuity problem in OBB regression, we propose to detect SAR ships by learning polar encodings. The encoding scheme uses a group of vectors pointing from the center of the ship target to the boundary points to represent an OBB. The boundary discontinuity problem is avoided by training and inference directly according to the polar encodings. In addition, we propose an Intersect over Union (IOU)-weighted regression loss, which further guides the training of polar encodings through the IOU metric and improves the detection performance. Experiments on the Rotating SAR Ship Detection Dataset (RSSDD) show that the proposed method can achieve better detection performance over other comparison algorithms and other OBB encoding schemes, demonstrating the effectiveness of our method.
Introduction: The Challenge of Oriented Ship Detection
Detecting ships in Synthetic Aperture Radar (SAR) images is vital for maritime safety and environmental monitoring. However, ships are rarely aligned with the axes of an image; they are slender and can point in any direction.
Standard object detection typically utilizes Horizontal Bounding Boxes (HBB). For a tilted ship, an HBB includes excessive "empty" sea or land clutter and often overlaps with neighboring ships in dense ports. To resolve this, researchers have turned to Oriented Bounding Boxes (OBB).
However, OBBs are mathematically complex to regress. When a model predicts the angle of a ship, it encounters the boundary discontinuity problem. For example, an angle of $89^\circ$ and $-91^\circ$ might represent nearly the same physical orientation, but to a neural network's loss function, they are numerically distant. This "angle flip" causes training instability and performance degradation.
Proposed Method: Polar Encodings
This research moves away from regressing angles entirely, proposing a Polar Encoding scheme that represents ship targets as a collection of vectors radiating from a center point. This transforms a discontinuous angular regression problem into a continuous distance regression problem.
Key Components of the Architecture:
- Feature Extraction: The model utilizes a ResNet-101 backbone paired with a Feature Refinement Network that fuses high-resolution shallow features with semantically rich deep features.
- Three-Branch Detection Head: The network predicts a center heatmap (localization), a center offset (quantization correction), and the polar encoding (the lengths of $N$ vectors defining the boundary).
- IoU-Weighted Loss: The authors introduced a regression loss weighted by Intersection over Union (IoU). This ensures the model prioritizes accuracy on the vectors that most impact the final shape's overlap with the ground truth, effectively forcing the network to focus on the ship's long axis.
- Decoding: During inference, predicted distances are converted back into boundary points, and a "Minimum Bounding Box" algorithm calculates the final oriented rectangle.
Experimental Results
The model was evaluated on the Rotating SAR Ship Detection Dataset (RSSDD), which contains 1,160 SAR images with 2,456 ship targets across various resolutions and polarizations. The proposed method outperformed existing angle-based and point-based OBB detection methods:
- Inshore Accuracy: On difficult, cluttered inshore datasets, the polar-based method achieved an Average Precision (AP) of 0.7942, significantly higher than the angle-based baseline of 0.6046.
- Overall Performance: On the full RSSDD set, it achieved an F1-score of 0.8594 for inshore and 0.9447 for offshore scenes.
- Efficiency: The model processes images in approximately 45ms, making it suitable for near-real-time maritime surveillance.
Conclusion
By rethinking ship geometry through polar encodings, this method elegantly solves the boundary discontinuity problem that has long plagued oriented object detection. The use of symmetry as a constraint and IoU-weighted loss leads to more accurate and stable detection of slender, tilted ships in complex SAR environments. Future work may explore adaptive sampling to better capture the geometry of ships with extremely high aspect ratios.
Summary updated