A Prompter Guided Distribution Matching for Zero-Shot SAR ATR

Zero-shot learning (ZSL) in SAR automatic target recognition (ATR) aims to recognize targets of the unseen classes. ZSL methods have been widely studied in optical images; however, most of them suffer from lacking shared semantics space and domain shift problems in SAR images. To address these, we propose an optimizable visual prompter (OVP) that automatically extracts semantic features from casually selected optical images, improving flexibility compared to strict alignment methods. We also introduce an optimal transport (OT)-based distribution matching method to align unseen SAR images with learned category-agnostic semantics, addressing domain shift issues. The OT method is theoretically optimal for matching samples from two distributions, making it particularly effective for zero-shot SAR ATR. Experimental results on real-world SAR datasets validate the efficiency and accuracy of our approach for small- and large-scale zero-shot tasks.