An Unsupervised Learning Framework With Efficient Weld Seam Extraction for Weld Defect Detection in Radiographic Images

Radiographic testing (RT) using X-rays is essential for ensuring weld integrity in shipbuilding. However, current RT inspection relies on manual evaluation, which is subjective, time-consuming, and costly. Therefore, automation of this inspection process is needed to improve efficiency and reliability. This automated approach involves two essential stages, extracting weld seams from raw radiographic images and detecting defects within them. Deep learning has emerged as a promising approach for this automation, but faces challenges from data heterogeneity in RT images and scarcity of weld defect data. Previous studies have addressed these stages separately or required large labeled datasets, thus lacking a complete pipeline for industrial deployment.We propose an end-to-end framework that addresses both stages with minimal data requirements. The framework first employs few-shot segmentation to extract weld seams from heterogeneous RT images using a fewlabeled examples. Subsequently, it detects weld defects through unsupervised learning trained on 183 normal weld images collected from a shipyard. Experimental results show 94.5% accuracy and 95.7% recall on shipyard data. Additionally, the framework achieves pixel-level anomaly localization without pixel-level annotations, enabling field deployment with minimal labeling burden.