SARCapsNet: A Capsule Network for High-Accuracy SAR Automatic Target Recognition Under Limited Training Sample Conditions

Most existing synthetic aperture radar (SAR) automatic target recognition (ATR) methods perform well with large-scale training data, but their performance significantly decreases in small-sample scenarios. Therefore, developing a SAR ATR method that can provide high accuracy and lightweight performance under different training sample sizes is a highly challenging problem. To address this problem, this paper pro poses a capsule network, named SARCapsNet, which consists of a hierarchical adaptive threshold capsule encoder (HATCE) and a reconstruction decoder. Particularly, HATCE is the most significant innovation of the proposed SARCapsNet. Specifically, in the proposed HATCE, an adaptive deformable threshold module (ADTM) is first developed by rationally fusing deformable convolution and soft threshold denoising techniques. This ADTM can adaptively capture the complex morphological features and key scattering distribution of SAR targets, while mitigating the impact of speckle noise. Second, different from ADTM, an adaptive thresholding module (ATM) is developed by reasonably integrating standard convolution with soft threshold denoising so that such ATM can further learn morphology-stable high-level features and enhance the model's ability to capture more robust and generalized structural features. Furthermore, a shuffle attention mechanism is incorporated into the HATCE to adaptively emphasize important features by capturing fine-grained, pixel level pairwise relationships and inter-channel dependencies, while suppressing irrelevant information. By ingeniously integrating ADTM, ATM, and shuffle attention mechanisms into HATCE, SARCapsNet can not only provide excellent recognition accuracy under different numbers of training samples but also exhibit lightweight characteristics. The experimental results show that SARCapsNet not only outperforms state-of-the-art models under both limited and full training sample conditions, but also maintains its lightweight characteristics. For instance, SARCapsNet achieves a recognition accuracy of 95.75% on the MSTAR dataset with only 10% of the training samples.