Class-Agnostic Dual-Branch Network for Few-Shot SAR Target Recognition

To address potential bottlenecks in cross-domain transfer for few-shot SAR target recognition, we overcome the limitations of source domain category attributes and propose a model named Class-Agnostic Dual-branch ProtoNet (CADP-Net). In this network, the global and local branches are designed to extract global and local features of the image, respectively, which are then mapped to a more discriminative space through a multi-layer perceptron. The dual-branch network is trained using moving exponential averaging under the consistency constraint of classification prediction. This approach can generalize a universal recognition strategy regardless of the target's category attributes, reduce intra-class distances, and increase inter-class distances. As a result, it enhances the ability to learn transferable representations in cross-domain few-shot tasks for SAR images. Experiments on the MSTAR dataset effectively validate our method.