Generalized Few-shot Anomaly Segmentation Framework for Industrial Surface Defect Detection

In industrial manufacturing, surface defect detection is essential for quality assurance. However, practical applications are often hindered by scarce defect samples, imbalanced categories, and the presence of unforeseen defect types, which limit the effectiveness of conventional fully-supervised and few-shot learning methods. Existing generalized few-shot segmentation approaches can handle multiple known classes but struggle to recognize unknown targets. To address this challenge, we propose AD-GFS-Seg (Anomaly Detection based General Few-shot Defect Segmentation), a novel framework that integrates an explicit anomaly-aware mechanism into few-shot segmentation. During training, the model learns the feature distribution of normal surfaces from defect-free samples. It is then fine-tuned using base-class and few-shot novel-class defect samples to construct class-specific prototypes. At inference, anomaly detection first identifies all anomalous regions, followed by prototype-based classification to achieve precise multi-class labeling. Experiments on the MVTec AD dataset demonstrate that AD-GFS-Seg outperforms state-of-the-art methods in known defect segmentation for most categories and significantly improves unknown defect segmentation. This work advances generalized few-shot segmentation in complex industrial environments and offers an effective solution to the open-world problem of recognizing unknown defect classes.