Transductive Prototypical Attention Network for Few-shot SAR Target Recognition

In recent years, synthetic aperture radar (SAR) automatic target recognition (ATR) methods driven by huge training samples have achieved remarkable results. However, in real SAR application scenarios, it is extremely difficult to provide enough training samples. This paper proposes a novel method, named transductive prototypical attention network (TPAN), to solve the few-shot target recognition problem in SAR ATR. The proposed method consists of three parts in total, i.e., region awareness-based feature extraction model, cross-feature spatial attention module and transductive prototype reasoning algorithm. Specifically, we build a region awareness-based feature extraction model, which can effectively focus on target regions of interest by embedding direction-aware and position-sensitive information. Next, a cross-feature spatial attention module is used to enhance the discriminativeness of identifying sample features. Finally, we propose a novel class inference algorithm, named transductive prototype reasoning algorithm, which iteratively updates class prototypes by mixing training samples with high-confidence test samples to achieve better class representation ability. Experimental results on moving and stationary target acquisition and recognition (MSTAR) dataset show that TPAN is effective and superior to some state-of-the-arts methods in few-shot SAR target recognition tasks.