Refined Prototype Classification Network with Multi-Dimension Feature Embedding for Few-Shot SAR Target Recognition

In recent years, the rapid development of deep learning has significantly advanced the automatic target recognition (ATR) in synthetic aperture radar (SAR) images. Although existing deep learning-based SAR ATR methods perform promising results, they are confronted with the issue of requiring a substantial amount of labeled samples for training. However, in the real application scenarios, acquiring a large number of samples is an extremely costly task, even in many cases, it may be impossible to accomplish. Therefore, it is extremely important to robustly achieve SAR target recognition under conditions of sample scarcity. In this paper, we propose a method called refined prototype classification network (RPCNet) with multidimensional feature embedding to address SAR target recognition in the case of limited samples. Specifically, we first propose a multi-dimensional feature embedding network, which is devised to sequentially extract the shape and multi-scale information of the target, so as to extract multi-dimension discriminative features from SAR imagery. Then, we introduce an efficient feature refinement module based on attention mechanism to highlight useful information and suppress redundant information, thus enhancing the discriminability of the features. Finally, a refined prototype classifier is proposed to achieve robust few-shot SAR target recognition with the help of transductive learning and inductive learning. Extensive experiments on the moving and stationary target acquisition and recognition (MSTAR) benchmark dataset demonstrate that the proposed method outperforms some of the state-of-the-art approaches.