Despeckling Representation for Few-Shot SAR Target Recognition

Due to the limited availability of labeled SAR data, few-shot SAR target recognition has become an active area of research. To address the inadequate performance of conventional deep learning models in this context, we propose DespeckleREP, a novel method that leverages despeckling-guided representation learning for few-shot SAR target recognition. Specifically, DespeckleREP adopts Restormer as the backbone network and introduces self-supervised SAR image despeckling as a pretext task to facilitate representation learning. This design provides two key benefits: (1) as a low-level vision task, SAR despeckling guides the network to extract essential low-level visual features of targets, such as edges and structural details, which are critical for recognition; and (2) the self-supervised nature of the task enables pretraining on large volumes of unlabeled SAR images, improving the generalization capability of the learned representations. After pretraining, the Restormer is jointly fine-tuned with a classification head for the target recognition task. Experimental results on a dataset comprising 20 target categories show that DespeckleREP achieves $72.58 \%$ recognition accuracy under the 10 -shot setting, significantly outperforming the baseline model, which achieves only $\mathbf{4 1. 8 4 \%}$ accuracy.