A self-configuring transformer segmentation method for welding radiographic defect detection in steel pipes

Radiography is a primary non-destructive testing method and is crucial for ensuring the quality of steel pipe welds. Although deep learning-based methods have shown significant achievements, they still encounter subjective biases during the network configuration. In addition, the problems of quantitative segmentation of small defects and motion blur remains challenging. To address these issues, this study constructed a pipe weld defect detection system based on CCD camera and image intensifier. Defect data are systematically collected and enhanced to establish a comprehensive pixel-level dataset of steel pipe weld defects. On this basis, a self-configuring transformer-based deep network (SCTnet) is proposed. It is capable of self-configuring parameters for any new dataset to generate a segmentation pipeline, reducing the subjectivity of the network. A multi-scale transformer module is proposed to integrate discriminative features across various semantic levels from a global perspective, effectively addressing the challenge of detecting small defects. Additionally, a hybrid loss function is proposed to effectively reduce the impact of motion blur. The results of using SCTnet on the steel pipe welds dataset demonstrate that the proposed segmentation method achieves competitive outcomes compared to other segmentation algorithms, with an average accuracy of 99.55% and an average dice coefficient of 78.91%.