Global-Local Information Interactive Learning Network for SAR Target Recognition with Limited Samples

Deep learning-based synthetic aperture radar (SAR) automatic target recognition (ATR) have shown great potential recently, however, the performance of these methods is subject to the number of annotated samples. In real application scenarios, it tends to acquire limited number of samples due to the acquisition cost, in which case the exising ATR method is susceptible to over-fitting. To achieve SAR target recognition robustly in the case of limited samples, this paper proposes a global-local information interactive learning network. Specifically, we first develop a global-local interactive learning architecture, which is dedicated to extract global-local discriminative feature by interactively integrating the merits of local convolution and sparse self-attention. Then, a hierarchical feature discrimination module is proposed to improve intra-class compactness and inter-class divergence, thereby boosting the recognition performance of the ATR model. Evaluation experiments on the moving and stationary target acquisition and recognition (MSTAR) dataset illustrate that the proposed method is superior to advanced SAR ATR methods under the condition of limited samples.