Research on the recognition of welding defects by modified ViT model combined with Canny algorithm

An improved Vision Transformer (ViT) model incorporating the Canny algorithm is proposed to address the low accuracy and efficiency issues in manually interpreting radiographic images for weld defect detection. The aim is to assist in recognizing typical welding defects and thereby enhance the accuracy and efficiency of interpretation. To enhance the model's capability in capturing local information, an edge feature extraction module based on the Canny algorithm is incorporated into the original ViT architecture, leveraging the advantage of global perception in ViT while addressing its limitations in handling local texture features. By parallelizing the edge extraction module with the ViT main architecture and fusing features at appropriate positions, the processing of global and local information is effectively balanced, leading to improved model generalization. Transfer learning is employed to reduce the model's data requirements during training, addressing the challenge of acquiring large-scale, high-quality weld radiographic datasets. The research findings demonstrate that the proposed model achieves significant performance improvements in weld defect recognition, achieving recognition accuracies of 98.97% and 95.94% on self-built datasets and the RIAWELC dataset, respectively. Compared to the original ViT model, these results represent an increase of 2.24% and 0.27%, respectively. Additionally, when compared with other mainstream image recognition models, enhancements ranging from 0.74% to 4.12% are observed with the proposed model, thus meeting the practical demand for interpreting weld radiographic images in production scenarios.