Abstract: Huge imbalance of different scenes' sample numbers seriously reduces Synthetic Aperture Radar (SAR) ship detection accuracy. Thus, to solve this problem, this letter proposes a Balance Scene Learning Mechanism (BSLM) for offshore and inshore ship detection in SAR images.
The Challenge of Scene Imbalance
Synthetic Aperture Radar (SAR) is a critical tool for maritime surveillance, yet ship detection models frequently underperform in coastal (inshore) environments. The primary obstacle is scene imbalance. In most standard datasets, offshore images—where ships appear as distinct, bright signatures against a uniform, dark water background—vastly outnumber inshore images.
Because offshore scenes are visually simple, deep learning models quickly master them. However, inshore scenes are fraught with complex land clutter, docks, and coastal infrastructure that mimic ship signatures. This disparity causes models to develop a "learning bias," over-optimizing for easy offshore cases while failing to accurately detect ships in high-stakes coastal areas, leading to high false-alarm rates.
The Balance Scene Learning Mechanism (BSLM)
To address this without the need for expensive manual labeling, the authors propose BSLM, a three-step unsupervised pipeline:
- Unsupervised Feature Extraction: A Generative Adversarial Network (GAN) is employed to learn the underlying distribution of SAR scenes. By concatenating feature maps from the discriminator's final layers, the system generates a robust "scene feature vector" for every image.
- Scene Clustering: These vectors are processed via K-means clustering to automatically group images. Given the statistical nature of SAR data, the smaller cluster is identified as "inshore" and the larger as "offshore."
- Balanced Augmentation: To eliminate learning bias, the minority inshore samples are augmented through replication, rotation, or Gaussian noise until the dataset achieves a balance between the two scene types.
This balanced data is then used to train standard detectors, including Faster R-CNN, Cascade R-CNN, SSD, and RetinaNet.
Key Results and Insights
The BSLM was validated using the SSDD (SAR Ship Detection Dataset), demonstrating consistent performance gains across all tested architectures:
- Significant mAP Improvements: The mechanism improved overall mean Average Precision (mAP) for Faster R-CNN (+2.87%), Cascade R-CNN (+2.86%), SSD (+7.66%), and RetinaNet (+6.34%).
- Inshore Performance Jumps: The most dramatic impact was seen in complex scenes. Specifically, SSD’s inshore mAP increased by 18.15%, and RetinaNet’s by 12.40%.
- Architectural Sensitivity: One-stage detectors (SSD, RetinaNet) showed much larger gains than two-stage detectors, suggesting that one-stage architectures are particularly vulnerable to background complexity and benefit most from scene balancing.
- Efficiency: As a training-time mechanism, BSLM enhances accuracy without compromising the inference speed of the final deployed model.
Conclusion
BSLM effectively mitigates the "over-learning" of offshore samples by forcing networks to spend more training iterations on complex inshore data. While an accuracy gap between inshore and offshore detection remains, BSLM significantly narrows it, providing a universal, unsupervised method to improve SAR maritime surveillance in challenging coastal environments.
Summary updated