A Distributed and Trusted Collaborative Framework for Industrial Defect Detection Based on Pseudo-Anomaly Augmentation and Residual Segmentation

As intelligent manufacturing continues to advance, automated surface defect detection has become critical to improving product quality and production efficiency. However, industrial image data are often characterized by sample scarcity, complex defect patterns, and data isolation across enterprises, which severely limits the effectiveness and generalization ability of defect detection models. To address the challenge of privacy-preserving and trustworthy collaborative detection, this paper proposes a distributed and trusted collaborative framework for industrial defect detection. Specifically: (1) a pseudo-defect generation strategy based on controllable perturbation and Perlin noise is introduced to enhance anomaly diversity and training robustness; (2) a reconstruction–residual–segmentation coupled model is designed to enable pixel-level localization of complex defects; (3) a federated learning framework with differential privacy is developed, in which blockchain-based smart contracts ensure verifiable and automated parameter aggregation under data-isolated conditions. Extensive experiments on the MVTec AD dataset demonstrate that the proposed method achieves superior performance across various categories, with average AP/AUROC scores of 80.8%/98.8% on texture classes and 79.9%/98.8% on object classes, significantly outperforming existing state-of-the-art methods.