Cross-domain few-shot defect recognition for metal surfaces

Abstract

Abstract Defect recognition for metal surfaces in the industry has attracted more and more attention. However, defect data scarcity presents a huge challenge for defect recognition in real industrial scenarios. The traditional few-shot defect recognition method can address this problem when the training data and test data are collected from the same or a similar metal surface. However, the defect data from similar metal surfaces are difficult to acquire to a certain extent. In this paper, we introduce a novel task setting that can achieve few-shot defect recognition by transferring knowledge across domains. The method consists of two levels: image-level and feature-level. At the image-level, a meta-augmentation method is proposed to improve the recognition generalization in each meta-task by joint parameter updating from the original and augmented domains. At the feature-level, a class covariance-guided feature perturbation method is proposed to perturb the feature distribution to enhance the cross-domain generalization capability. The extension of cross-domain experiments from textured to metal surfaces shows the superior performance of the proposed method compared to other mainstream methods.