Inverse Synthetic Aperture Radar Space Target Recognition Algorithm Based on Few-Shot Class-Incremental Learning

Abstract

Identifying the space targets rapidly and accurately is a significant mission in defending national rights and maintaining space security. The continuous emergence of new space targets has led to a dramatic decrease of the performance of Inverse Synthetic Aperture Radar (ISAR) automatic target recognition. This paper proposes a few-shot space target recognition method, which can adaptively excavate the new class features to improve the ISAR target recognition performance. Algorithm proposed in this paper facilitates existing data to predict unseen class and constructs the corresponding feature embedding space, which build the basis for unseen class recognition. Algorithm using memorizing complementary networks to extract global-focus features and local-focus features, which is conductive to increasing recognition accuracy while enhancing feature richness. To address the problem of catastrophic forgetting, this paper proposes an instance mining method based on inductive probability to extract hard-discriminative features at a deep level. Combing with these improving components mentioned above, algorithm of this paper can achieve compatibility and balance between new and old categories effectively. The experimental results demonstrate that algorithm proposed in this paper is superior to other advanced methods in the task of incremental space target recognition.