FSD³: Few-Shot SAR Object Detection With Dynamic Perception RPN and Category Knowledge Decoder

Limited by observational constraints and non-cooperative targets, SAR images often suffer from extreme scarcity of training samples for certain classes, which severely degrades the detection performance and leads to the challenging issue of few-shot SAR object detection. To address this, we propose a Few-Shot SAR object Detection model with Dynamic perception RPN and category knowledge Decoder, dubbed as FSD<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>. Specifically, a dynamic perception RPN is employed to infuse global semantics of support classes into local features of query feature maps, thereby enhancing the recall rate of novel-class targets under few-shot conditions. Subsequently, a category knowledge decoder is designed, which automatically decodes aggregated features by utilizing query features as prompts and effectively mitigates the feature-label mismatch phenomenon. Finally, a dual-constrained detection head is proposed, which leverages the large margin cosine loss and center calibration loss to explicitly achieve inter-class separation and intra-class compactness, thereby achieving high-precision and robust classification in few-shot scenarios. Experimental results on the public SAR-AIRCraft-1.0 dataset and our custom MSTAR-Fewshot dataset have demonstrated that the proposed model achieves superior few-shot object detection performance across various experimental configurations.