Open-Set SAR Target Recognition Based on Observation-Invariant Contrastive Learning

Open-set recognition (OSR) focuses on accurately recognizing known class samples while identifying unknown ones. Existing methods typically enhance the distinction between known and unknown classes by optimizing feature representations. However, compared to general OSR tasks, the open-set recognition of synthetic aperture radar (SAR) targets poses greater challenges. First, SAR target recognition is highly sensitive to variations in observation conditions, leading to unstable feature extraction and degraded generalization across different observation conditions. Second, the acquisition and annotation of SAR data are costly and labor-intensive, resulting in limited training samples and increased difficulty in learning reliable decision boundaries. To address these challenges, we propose adaptive triplet center loss (ATCL) with observation-invariant contrastive learning (OICL) for open-set SAR target recognition. In OICL, variations in observation conditions are explicitly simulated through data transformations, encouraging the model to learn stable and consistent representations despite these changes. Consequently, the adverse effects of observation sensitivity are effectively mitigated. Then, the intra-class compactness and inter-class separability of feature representations are enhanced through ATCL. In ATCL, the difficulty of each sample is adaptively assessed, and higher optimization priorities are given to hard samples, enabling a more thorough refinement of feature representations. During testing, the maximum logit score (MLS) is employed as the decision criterion instead of the commonly used maximum softmax probability (MSP), reducing overconfidence in predictions and thereby improving the distinction of unknown classes. Extensive experiments conducted on the SAR dataset demonstrate that the proposed method achieves state-of-the-art performance.