Pseudo-Unknown Class Guided-Based Open-Set Learning Network for SAR Automatic Target Recognition

Most synthetic aperture radar (SAR) automatic target recognition (ATR) methods are developed for the closed-set environment, so these ATR methods can only identify known classes in the target library. However, it is known that the real scenario is open, so it requires the ATR model should be capable of identifying unknown categories while classifying known categories. Therefore, this paper proposes a pseudo-unknown class guided-based open-set learning network for SAR ATR tasks. First, to enhance the separability between known categories, a separable feature embedding space based on von Mises-Fisher (vMF) distribution is established. Second, unknown class decision boundaries are constructed based on pseudo-unknown classes synthesized by known categories, and are expanded based on the idea of contrastive learning, which is very helpful in promoting the ability of the model to identify unknown categories. Experiments on moving and stationary target acquisition and recognition (MSTAR) dataset demonstrate the effectiveness of the proposed method.