Expert-aware multi-stage AI-assisted image processing pipeline for accurate weld defect detection from radiographs: a use case for metal fabrication industries

Radiography-based non-destructive testing is vital in shipbuilding, automotive, and aerospace industries, but manual inspection is labour-intensive and error-prone. While deep learning offers automation, it often lacks explain ability and expert-level insight. This paper presents a multi-stage, expert-aware AI pipeline for automated weld defect detection in radiographic images. The system targets five common defects: blowhole, porosity, wormhole, lack of fusion, and inclusion. Radiographic 16-bit DICOM images are enhanced and converted to 8-bit PNGs using histogram-based contrast adjustment guided by domain thresholds. Weld zones are detected using a YOLO v11 model (97.2% accuracy) and divided into square clips. Each clip undergoes binary classification via transfer learning (F1 score: 97%) to distinguish defective from non-defective regions. Defective clips are further analysed using a multi-label classifier to identify specific defect types. A secondary YOLO v11 model (93.1% accuracy) pinpoints defect locations, which are clustered and evaluated using ISO/NES standards. The pipeline integrates rules-based logic, simulated visual attention, and supervised learning to mirror expert inspection practices. It achieves 96% accuracy compared to human weld supervisors, offering a reliable, interpretable, and domain-informed solution for automated weld quality assessment.