Deep transfer learning + BDENet: An innovative approach for detecting the internal defects of weld seams using radiographic imaging

Abstract

The radiographic testing of weld defects is crucial for ensuring welding quality. Identifying these defects is challenging because radiographic images feature low brightness and contrast. The traditional measurement and grading of defect sizes heavily relies on the subjective judgment and operational proficiency of non-destructive testing personnel, which compromises the accuracy and consistency of results. This study proposes an automatic weld defect detection method based on deep transfer learning and BDENet to achieve the precise identification of various weld defects. Firstly, the backbone network was pre-trained on Pascal VOC2012 and SBD datasets. The optimal weights were determined by fine-tuning hyperparameters, which enhanced the feature extraction ability of the model. Subsequently, the datasets were processed using overlapping clipping, which enhanced the datasets and retained more boundary information. Then a novel BDENet model was introduced. The model featured an boundary detail enhancement network (BDENet) consisting of detail enhancement convolution and differential convolution. BDE enhanced the semantic information of boundary details in weld images by integrating prior information and differential features through residual addition. This enhancement significantly improved segmentation accuracy and sensitivity to small-scale features. Finally, the fine-tuned transfer learning technology was applied to the BDENet model. The proposed method outperformed other semantic segmentation models, achieving remarkable metrics: 86.70 % mIoU, 91.17 % mPA, 92.71 % mF1, and 90.99 % mR. These results validate the model's effectiveness, demonstrating its superiority in the semantic segmentation of radiographic images of defects.