Eternal-MAML: a meta-learning framework for cross-domain defect recognition.

Defect recognition tasks for industrial product suffer from a serious lack of samples, greatly limiting the generalizability of deep learning models. Addressing the imbalance of defective samples often involves leveraging pre-trained models for transfer learning. However, when these models, pre-trained on natural image datasets, are transferred to pixel-level defect recognition tasks, they frequently suffer from overfitting due to data scarcity. Furthermore, significant variations in the morphology, texture, and underlying causes of defects across different industrial products often lead to a degradation in performance, or even complete failure, when directly transferring a defect classification model trained on one type of product to another. The Model-Agnostic Meta-Learning (MAML) framework can learn a general representation of defects from multiple industrial defect recognition tasks and build a foundational model. Despite lacking sufficient training data, the MAML framework can still achieve effective knowledge transfer among cross-domain tasks. We noticed there exists serious label arrangement issues in MAML because of the random selection of recognition tasks, which seriously affects the performance of MAML model during both training and testing phase. This article proposes a novel MAML framework, termed as Eternal-MAML, which guides the update of the classifier module by learning a meta-vector that shares commonality across batch tasks in the inner loop, and addresses the overfitting phenomenon caused by label arrangement issues in testing phase for vanilla MAML. Additionally, the feature extractor in this framework combines the advantages of the Squeeze-and-Excitation module and Residual block to enhance training stability and improve the generalization accuracy of model transfer with the learned initialization parameters. In the simulation experiments, several datasets are applied to verified the cross-domain meta-learning performance of the proposed Eternal-MAML framework. The experimental results show that the proposed framework outperforms the state-of-the-art baselines in terms of average normalized accuracy. Finally, the ablation studies are conducted to examine how the primary components of the framework affect its overall performance. Code is available at https://github.com/zhg-SZPT/Eternal-MAML.