Aiming at the problem of weld defect images affected by high noise and interference fringes in TOFD (time of flight diffraction) ultrasonic detection technology, as well as the challenge of feature information loss and computational efficiency imbalance faced by current deep learning models in processing such images, an innovative defect identification model integrating convolutional neural network (CNN) and Transformer architecture is proposed, named MHLFNet (multi-scale high-low focused network). By introducing a multi-scale feature fusion (MSFF) module, this model significantly enhanced the ability of capturing local information. At the same time, a high-low focused linear (HLFL) module was designed, which used the adjustable allocation ratio to dynamically allocate the attention for the high and low frequency information of feature images, and adopted focused linear attention instead of traditional multi-head self-attention, effectively reducing the computational complexity while enhancing the diversity of attention mechanisms and the feature expression ability. In order to verify the performance of MHLFNet, a TOFD weld defect image dataset was constructed, and a systematic experimental evaluation was conducted. The results showed that MHLFNet achieved an accuracy of 98.6% in the weld defect identification task, and performed excellently in terms of model parameters, floating-point operations, and inference time. In visual analysis and identification validation, MHLFNet demonstrates excellent identification capabilities for high-risk defects (such as lack of fusion and cracks), proving its reliability and engineering value in industrial inspection.