A Multiple Level Unsupervised Domain Adaption Network for SAR Automatic Target Recognition

In recent years, Synthetic Aperture Radar (SAR) automatic target recognition (ATR) has made significant progress alongside the rapid development of deep learning. Typically, deep learning-based SAR-ATR methods require a large amount of labeled data to train models, which presents a major challenge, as annotating SAR images is extremely costly. Using simulated samples to train deep models offers a viable alternative. However, due to the significant distribution gap between the simulated and real domains, cross-domain recognition remains challenging. In this paper, we propose a multiple level unsupervised domain adaptation (UDA) network for SAR-ATR that mitigates the domain shift between the source and target domains. Specifically, we align the distribution discrepancies between simulated and real samples at three different scales: pixel-level, domain-level, and class-level, which enhances the transferability of the source domain network. In addition, a pseudo-label filtering strategy based on contrastive learning is employed to improve the quality of pseudo-labels and further enhance class-level alignment. Finally, we conduct extensive experiments on two datasets. The results demonstrate that the proposed network effectively reduces the distribution gap between the simulated and real domains, learns useful information from the simulated samples, and successfully transfers this knowledge to the real domain. This approach significantly decreases the reliance on large quantities of labeled samples in real-world SAR-ATR applications.