Learning Polar Encodings for Arbitrary-Oriented Ship Detection in SAR Images

Yishan He, Fei Gao, Jun Wang, Amir Hussain, Erfu Yang, Huiyu Zhou

Abstract

Common horizontal bounding box-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 article, 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. Comparative experiments on the benchmark Rotating SAR Ship Detection Dataset (RSSDD) demonstrate the effectiveness of our proposed method in terms of enhanced detection performance over state-of-the-art algorithms and other OBB encoding schemes.

Introduction: The Challenge of Oriented Ship Detection

Detecting ships in Synthetic Aperture Radar (SAR) images is a critical task for maritime security and traffic monitoring. Unlike optical sensors, SAR can operate through cloud cover and at night. However, ships in SAR imagery are typically slender and appear at arbitrary orientations.

Traditional object detection methods rely on Horizontal Bounding Boxes (HBB). While effective for many objects, HBBs are ill-suited for slender, diagonal ships because they encompass excessive background clutter (sea or land), leading to high overlap and false positives in dense environments like ports. To address this, researchers have turned to Oriented Bounding Boxes (OBB).

Oriented Bounding Boxes (OBB) provide a much tighter fit for slender ship targets compared to standard horizontal boxes.

Despite their advantages, regressing OBBs introduces the "boundary discontinuity" problem. Most models predict an angle ($\theta$), but angles are periodic; for instance, $0^\circ$ and $180^\circ$ represent the same physical orientation but are numerically distant. This causes instability in the loss function during training, hindering performance.

Methodology: Polar Encoding and IoU-Weighted Loss

This paper proposes a novel approach that bypasses angle regression entirely by learning polar encodings.

Polar Encoding & Decoding

Instead of predicting width, height, and angle, the proposed model predicts the distance from the ship's center to its boundary at $N$ fixed angular intervals (the authors utilized $N=8$). These distances vary smoothly as the ship rotates, eliminating the numerical "jumps" associated with angular boundaries. During inference, the model reconstructs the boundary points from these distances and calculates the Minimum Bounding Box (MBB) for the final detection.

The proposed network architecture features specific branches for generating heatmaps, center offsets, and polar encodings.

IoU-Weighted Regression Loss

The authors also introduce an Intersection over Union (IoU)-weighted regression loss. Standard regression losses (like Smooth-L1) treat all coordinate errors equally. However, for a slender ship, an error in the "width" direction impacts the IoU much more severely than an error in the "length" direction. This new loss function weights errors based on their geometric significance, guiding the model to prioritize accuracy where it matters most for the IoU metric.

The underlying architecture utilizes a ResNet-101 backbone with a feature refinement network to fuse multi-scale information, ensuring the model can detect both small vessels and large tankers.

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 diverse "inshore" (high clutter) and "offshore" (open sea) environments.

  • Detection Performance: The method achieved an F1-score of 0.8594 in complex inshore environments and 0.9447 in offshore scenes, outperforming state-of-the-art models like BBAVectors and ROITransformer.
  • Efficiency: The model processes images in approximately 45ms, making it suitable for near real-time applications.
  • Robustness: The polar encoding scheme proved particularly effective in high-clutter inshore areas where land clutter often confuses traditional detectors.

Conclusion

By representing OBBs through polar encodings rather than direct angle regression, this work successfully mitigates the boundary discontinuity problem in SAR ship detection. While the method is highly effective, future work may involve moving the Minimum Bounding Box calculation to the GPU to eliminate CPU bottlenecks and exploring higher sampling densities ($N$) for more complex target geometries.

Summary updated

Abstract

Common horizontal bounding box-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 article, 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. Comparative experiments on the benchmark Rotating SAR Ship Detection Dataset (RSSDD) demonstrate the effectiveness of our proposed method in terms of enhanced detection performance over state-of-the-art algorithms and other OBB encoding schemes.
BibTeX
@article{he2021learning,
  author = {Yishan He and Fei Gao and Jun Wang and Amir Hussain and Erfu Yang and Huiyu Zhou},
  title = {{Learning Polar Encodings for Arbitrary-Oriented Ship Detection in SAR Images}},
  journal = {IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens.},
  year = {2021},
  url = {https://openreview.net/forum?id=pbsS0f0x7e},
}
Figure