R 2 -Net: A Hybrid Meta-Learning Framework for Cross-Domain Few-Shot Industrial Defect Detection

Few-shot industrial defect detection is critically challenged by poor cross-domain generalization, where models often fail to adapt from a source domain to new target domains. Existing meta-learning paradigms also face significant constraints in addressing this. Metric-based methods are prone to overfitting, while optimization-based approaches often suffer from high computational costs and training instability. To this end, we propose R 2 -Net, a hybrid meta-learning framework that balances performance and efficiency. Its core contribution lies in resolving the aforementioned dilemma through a synergy of optimization and inference: we employ the efficient first-order meta-optimizer Reptile to learn a high-quality set of meta-initial parameters. Building on this foundation, the model utilizes a backbone network integrated with an attention mechanism to extract high-quality features, which are then fed into a relation network for rapid and fine-grained relational defect inference. Experimental results on the MVTec AD and NEU-CLS datasets demonstrate that our framework significantly outperforms a range of strong baseline models in cross-domain few-shot tasks, validating its effectiveness.