Collaborative Learning of Visual-semantic Features for Few-shot SAR Target Recognition

To address the susceptibility of Synthetic Aperture Radar (SAR) target recognition models to overfitting and their limited inter-class discriminability under few-shot conditions, this paper proposes a few-shot SAR target recognition method that employs collaborative learning of visual and semantic features. First, given the limited visual information available in SAR imagery under few-shot conditions, a semantic information construction method for SAR targets is developed based on a “generation-encoding-selection” paradigm. Class-level semantic prototypes are then constructed from semantic prior information to capture target-class attributes and complement visual features. Second, a dual-branch collaborative learning network for visual and semantic features is developed to extract more discriminative class representations of SAR targets. In the visual branch, the Pearson correlation matrices of the visual and semantic prototypes are aligned, to transfer category-related information from the semantic prior information to the visual branch. A dynamic quadruplet loss is employed to optimize the network and enhance the discriminability of the visual features. In the semantic branch, optimal transport theory is introduced to refine semantic features using visual features, thereby improving their discriminative capability. Third, a weighted decision-level fusion strategy based on instance-level prediction confidence is proposed to achieve robust SAR target recognition by jointly exploiting semantic and visual features. Experiments were conducted on the standard and extended sets of the MSTAR dataset, as well as on the OpenSARShip dataset. The experimental results demonstrate that the proposed method effectively exploits the complementary strengths of visual-semantic target features under few-shot conditions, leading to substantial improvements in both recognition accuracy and robustness.