Automated Welding Defect Detection using Point-Rend ResUNet

In the field of welding inspection, radiographic non-destructive evaluation (NDE) is a widely used technique for detecting defects in welds. However, this technique requires professionally qualified workers to manually judge radiographs to determine the presence, type, and location of defects. Recently, deep learning techniques have been developed to automate this process by using image segmentation. Despite its effectiveness, small-size targets in segmentation can have blurred boundaries, making it difficult to accurately annotate them at the pixel level. In this study, we propose an automated approach using the Point-REND Res-UNet model to improve the accuracy of detecting welding defects. Our method uses the improved Point-Rend algorithm to iteratively refine coarse segmentation results, allowing for more accurate defect detection. We evaluate our approach on a set of X-ray data and demonstrate that it achieves an improvement in model dice of 6.22%. Our proposed approach can potentially save labor time and costs while enhancing the accuracy and efficiency of welding defect detection.