Class Ambiguity Regularized Adversarial Domain Adaptation for Cross-resolution SAR ATR

Automatic target recognition (ATR) plays a critical role in synthetic aperture radar (SAR) applications. However, existing SAR ATR methods usually assume that the training and test SAR samples have the same imaging resolution but it is hard to satisfy in practice due to the complexity of the SAR imaging process, which results in the discrepancy between different resolution domains and degrades the performance of ATR method. To this end, we propose a class ambiguity regular-ized adversarial domain adaptation method for cross-resolution SAR ATR. First, the feature extractor and domain discriminator are designed with the idea of adversarial learning to eliminate the domain discrepancy. Then, the classifier is used to learn the class information of labeled SAR images, enabling the accurate identification on the labeled domain before adapted to unlabeled target domain. Finally, the class ambiguity regularization is in-troduced to minimize recognition error over unlabeled resolution domain. Extensive experimental results based on the moving and stationary target acquisition and recognition (MSTAR) dataset have shown the superiority of the proposed method on cross-resolution SAR ATR.