A two-stage contrastive learning paradigm for fine-grained SAR object detection with limited labeled samples

Fine-grained object detection in Synthetic Aperture Radar (SAR) imagery remains challenging due to speckle noise, low signal-to-noise ratios, and the high cost of pixel-level annotation. While contrastive self-supervised learning has shown promising results in general vision tasks, existing methods are still limited—many focus predominantly on classification rather than detection, or rely solely on backbone-level pretraining without effectively incorporating the multi-scale representations required by detection frameworks. To address these limitations, we propose TsCL, a two-stage contrastive learning paradigm tailored for SAR object detection with limited labeled data. The first stage employs Dense Contrastive Learning to pretrain the backbone network, producing spatially aligned feature representations. The second stage introduces an FPN-level contrastive objective that leverages high-level pyramid features for global representation learning. Furthermore, to enhance noise robustness to speckle noise, we incorporate a domain-specific pretext task, that simulates noise patterns as part of the augmentation strategy during pretraining. Comprehensive experiments on both the SAR-Aircraft 1.0 and FAIR-CSAR datasets show that TsCL consistently outperforms fully supervised baselines and state-of-the-art self-supervised methods, especially in label-scarce scenarios. These results demonstrate the practical efficacy and generalization ability of our approach for real-world SAR object detection applications.