Novel Category Discovery Without Forgetting for Automatic Target Recognition

In this article, we explore a cutting-edge concept known as class incremental learning (CIL) in novel category discovery for synthetic aperture radar (SAR) targets (CNTs). This innovative task involves the challenge of identifying categories within unlabeled datasets by utilizing a provided labeled dataset as reference. In contrast to the conventional category discover approaches, our method introduces novel categories without relying on old labeled classes and effectively mitigates the issue of catastrophic forgetting. Specifically, to reduce the bias of the established categories toward unknown ones, CNT extracts representational information via self-supervised learning, gleaned directly from the SAR data itself to facilitate generalization. To retain the model's competence in classifying previously acquired knowledge, we employ a dual strategy incorporating the rehearsal of base category feature prototypes and the application of knowledge distillation. Our methodology integrates multiview and pseudolabeling strategies. In addition, we introduce a novel approach that focuses on enhancing the discernibility of class spaces. This strategy primarily ensures distinct separation of the unlabeled classes from base class prototypes, and imposes stringent constraints on the internal relationships among individual samples and their corresponding perspectives. To the best of our knowledge, this is the first study on category discovery in the CIL scenario. The experimental results show that our method significantly improves the performance on SAR images compared to the previous optimal method, which indicates the effectiveness of our method.