Unsupervised steel pipe weld defect detection based on internal and external guided reverse distillation

Non-destructive testing based on radiographic images is the key to the quality control of welded steel pipes. However, high annotation costs and the difficulty in dealing with unknown defects limit supervised methods. Existing unsupervised radiographic defect detection methods for welds mostly rely on reconstruction errors or single-feature distribution modeling, and they are still prone to false and missed detections under complex backgrounds and low-contrast subtle defect scenarios. Therefore, an unsupervised radiographic defect detection method for steel pipe welds based on intra- and inter-guided reverse distillation was proposed. Based on the reverse distillation framework, an intra- and inter-guided mechanism was introduced as a learnable prototype. Structured constraints were applied to the feature distribution from two levels of intra-sample and inter-sample to enhance the difference between local anomalies and contexts, thereby improving the discrimination ability for subtle defects under complex backgrounds. Experimental results indicate that compared with mainstream unsupervised anomaly detection methods, the proposed method achieves better results, and it provides a solution with engineering application value for the automation and intelligence of radiographic inspection for steel pipe welds.