LDCL: Low-Confidence Discriminant Contrastive Learning for Small-Sample SAR ATR

Synthetic aperture radar (SAR) target image acquisition presents challenges and incurs high annotation costs. The emergence of self-supervised contrastive learning shows promise for SAR automatic target recognition (ATR) with limited data. However, SAR images suffer from poor discriminability and high sample similarity, hindering instance discrimination in contrastive learning. To address this, we propose low-confidence discriminant contrastive learning (LDCL), which integrates group-instance contrast and batch-mixed training for SAR ATR. LDCL consists of two branches: classical instance discrimination and group-instance discrimination. We refine the SAR-group instance discrimination loss function by incorporating distance calculations to guide feature vectors toward the nearest clusters, enhancing discrimination within the feature space. In addition, we introduce a batch image mixing training strategy to reduce confidence in SAR instance discrimination while preserving intraclass consistency. Experimental results on small sample MSTAR and FUSAR-Ship datasets demonstrate that LDCL outperforms traditional transfer learning and self-supervised learning (SSL) methods, achieving significantly higher recognition rates in SAR ATR tasks.