A causal semantic intervention framework for few-shot SAR-ATR

Few-shot synthetic aperture radar automatic target recognition (SAR-ATR) remains challenging due to limited labelled samples. Most existing few-shot SAR-ATR methods recognize targets merely from visual features, and their recognition process is easily affected by biases introduced by pretrained knowledge. To address this issue, we propose a causal semantic intervention framework for few-shot SAR-ATR. The recognition process is formulated using a structural causal model, in which pretrained knowledge is treated as a latent confounder that affects both visual feature extraction and class-level representation. To mitigate this confounding effect, we use class-level semantic priors derived from a vision-language model to complement visual prototypes, thereby instantiating semantic intervention and reducing the model’s reliance on biased pretrained knowledge. Specifically, we implement this design in a dual-branch framework. The visual branch employs LoRA-based adaptation of the pretrained image encoder and constructs class prototypes from limited support samples. The semantic branch uses multiple textual prompts to generate diverse class-level semantic priors through a frozen text encoder. To improve the reliability of semantic guidance, a margin-guided semantic aggregation strategy is proposed to assign query-adaptive weights to prompt-specific logits based on their confidence margins. The visual and semantic predictions are then combined through logit-level fusion. Experiments on three benchmark datasets show that the proposed model generally improves few-shot recognition accuracy and robustness compared with existing SAR-ATR approaches.