Weld TOFD defect classification method based on multi-scale CNN and cascaded focused attention

Aiming at the problems of high noise and interference fringes of weld defect images in TOFD detection technology, and the challenges of feature information loss and computational efficiency imbalance faced by current deep learning models in processing such images. We innovatively propose a defect identification model of hybrid CNN and Transformer architecture named MCFNet (Multi Cascaded Focused Network). The multi-scale feature fusion (MSFF) module is introduced to enhance the ability of local information extraction. At the same time, an efficient and fast transformer module (EFTM) has been designed. In this module, a cascaded group attention (CGA) mechanism is employed to segment feature graphs, and focused linear attention is utilized instead of the traditional multi-head self-attention (MHSA). This design aims to reduce computational complexity and enhance the diversity of attention mechanisms. In order to verify the performance of the model, we constructed a TOFD defect dataset STTOFD-DEF and conducted extensive experiments. The experimental results show that MCFNet achieves a high accuracy of 98.72 % on defect identification, while maintaining a Params of 10.21 M, Flops of 0.423G and inference time of 55.60 ms, and surpasses the existing classical networks in many key indicators. In visualization and identification performance verification, MCFNet demonstrated the highest accuracy in identifying the most dangerous welding defects, such as Lack of fusion and Crack, demonstrating its reliability and effectiveness in practical engineering applications.