A Contrastive Uncertainty Modeling for Radar HRRP Domain Generalization Target Recognition

Radar High Resolution Range Profile (HRRP) is widely recognized in Radar Automatic Target Recognition (RATR) systems. In practical scenarios, the radar data collection process can vary due to differences in polarization methods and operating frequency bands, leading to distribution discrepancies (domain shift) across different HRRP datasets. This is the first work that introducing domain generalization to address the domain shift problem in HRRP recognition. Besides, existing domain generalization algorithms for HRRP data fail to address the uncertain statistical differences caused by domain shifts, often resulting in over-fitting. To tackle this, we propose a Contrastive Uncertainty Modeling (CUM) approach for radar HRRP domain generalization in target recognition. This method adopts a feature-statistical perspective to model potential uncertainty, effectively capturing the complexities and variations across data domains. As a result, the model can learn domain-invariant features, improving its robustness against domain perturbations. Additionally, we introduce a Hybrid Supervised Loss (HSL) function, which incorporates homologous consistency constraints to encourage the model to learn compact intra-class and well-separated inter-class feature representations. Experimental results on a comprehensive HRRP dataset demonstrate that the proposed algorithm outperforms the latest related works in terms of generalization ability. Furthermore, the proposed framework offers the advantages of plug-and-play functionality and an end-to-end design.