R-CenterNet+: Anchor-Free Detector for Ship Detection in SAR Images

Yuhang Jiang, Wanwu Li, Lin Liu2021 · DOI

TL;DR

This paper introduces R-CenterNet+, an anchor-free deep learning model for detecting ships in Synthetic Aperture Radar (SAR) images. It addresses challenges like small ship sizes, complex backgrounds, and arbitrary ship orientations by incorporating attention mechanisms, a foreground enhancement module, and a rotation detection head, achieving high accuracy and real-time performance.

What Problem Does This Paper Address?

Detecting ships in SAR images is crucial for maritime surveillance, but it's a challenging task. SAR images have unique characteristics like speckle noise and geometric distortions, making interpretation difficult. Traditional methods struggle with:

  • Sparse Targets & Computational Cost: Ship targets are often sparse, leading anchor-based detectors to generate many unnecessary anchor boxes, increasing computation and hindering real-time detection.
  • Small Ship Detection: Detecting small ships in large SAR scenes is particularly difficult.
  • Complex Backgrounds: Terrestrial interference and similar-looking objects can lead to false positives and missed detections.
  • Orientation and Size Ambiguity: Ships have large aspect ratios and arbitrary orientations, making it hard for standard horizontal bounding boxes to accurately capture their shape and axial information.

Methods

The proposed R-CenterNet+ builds upon the anchor-free CenterNet architecture and introduces several key improvements:

  • CBAM-DLA34 Backbone: The authors integrate the Convolutional Block Attention Module (CBAM) into the DLA (Deep Layer Aggregation) backbone. CBAM enhances feature representation by adaptively optimizing channel and spatial attention, improving the model's focus on small ship targets.
  • Foreground Enhance Module (FEM): To combat complex backgrounds and reduce false alarms, FEM uses semantic segmentation principles. It predicts foreground areas to refine feature maps, helping to distinguish ships from similar-looking background elements like islands and reefs.
  • Rotation Detection Head: Instead of traditional horizontal bounding boxes, R-CenterNet+ uses rotating bounding boxes to accurately represent ship orientation, size, and axial information. This is achieved by adding an angle regression branch to the detection head.

Datasets

The paper evaluates R-CenterNet+ on two publicly available SAR ship detection datasets:

  • SSDD (SAR Ship Detection Dataset): Established in 2017, it contains images from various sensors, resolutions, and polarizations, covering inshore and offshore scenes. The authors also created "SSDD-inshore" and "SSDD-offshore" subsets.
  • AIR-SARShip Dataset: Introduced in 2019, this dataset features high-resolution, large-scale SAR images from the Gaofen-3 satellite, also covering diverse scenes. Similar to SSDD, inshore and offshore subsets were derived.

Key Insights and Ideas

  • Anchor-Free Advantage for SAR: The paper reinforces that anchor-free methods are well-suited for SAR ship detection due to the sparsity of targets, offering a better balance between accuracy and speed compared to anchor-based approaches.
  • Attention for Small Targets: The integration of CBAM significantly boosts the detection of small ships, a known weakness in many SAR detection models.
  • Semantic Segmentation for Robustness: Using FEM, inspired by semantic segmentation, proves effective in filtering out confusing background elements, leading to fewer false positives in complex scenes.
  • Rotation Detection is Key: Accurately capturing ship orientation with rotating bounding boxes is crucial for precise detection and better feature extraction, especially for elongated targets.

Key Results

  • High Accuracy: R-CenterNet+ achieved an Average Precision (AP) of 95.11% on the SSDD dataset and 84.89% on the AIR-SARShip dataset.
  • Real-time Performance: The model demonstrates a detection speed of 33 frames per second (FPS), making it suitable for real-time applications.
  • Superiority over Baselines: R-CenterNet+ outperformed traditional anchor-based (Faster-RCNN) and anchor-free (Centernet) methods on both datasets, particularly in challenging inshore scenes.
  • Module Effectiveness: Ablation studies showed that CBAM improved AP by 0.80% and FEM by 1.09% on the SSDD dataset, highlighting the contribution of each component.

Limitations and Future Work

  • Dataset Differences: The authors note that the AIR-SARShip dataset's performance was lower than SSDD due to significant differences between its training and testing sets, suggesting that improving model migration ability across different large-scale scenes is a future research direction.
  • Further FEM Improvement: The paper suggests continuing to improve FEM, potentially by using instance segmentation to mark foreground areas more precisely, aiming for more refined foreground enhancement.
  • Dense Ship Detection: While the model shows promise, detecting extremely densely arranged ships remains a challenge, which the authors aim to address in future work by further distinguishing individual ship targets.

Summary updated

BibTeX
@misc{jiang2021r,
  author = {Yuhang Jiang and Wanwu Li and Lin Liu},
  title = {{R-CenterNet+: Anchor-Free Detector for Ship Detection in SAR Images}},
  year = {2021},
  doi = {10.3390/s21175693},
}
Figure