Self-Supervised Learning for Industrial Anomaly Detection in Smart Manufacturing

Defect detection has become a crucial component of intelligent manufacturing, with anomaly detection methods being widely used in various industrial tasks due to their significant effectiveness and efficiency. Addressing the diversity and complexity of anomalies in the detection process, we propose a method for anomaly detection and localization based on synthetic anomalies. Our method employs multi-scale feature fusion and self-attention skip connections to enhance reconstruction accuracy and suppress secondary reconstruction of anomalous regions. By reconstructing sample features and incorporating feature reconstruction errors into the pixel-level anomaly segmentation network, we mitigate noise interference during the reconstruction process. Experiments on the MVTec AD dataset show our method achieves superior performance, with an image-level anomaly detection AUROC score of 99.4%, offering a robust solution for industrial anomaly detection in intelligent manufacturing.