Abstract: To address the problem of insufficient detection accuracy of visible-light vision–based methods in surface defect identification of welded joints, which is mainly caused by illumination variations, simple backgrounds, small defect sizes, and inconspicuous features, a weld surface defect detection method based on infrared imaging and deep learning is proposed. An infrared thermal imager is used to acquire temperature field information of aluminum alloy friction stir welded joints during the post-weld residual heat stage, and an infrared image defect dataset is established. Based on the YOLO model, classification and localization of weld surface defects are realized. To improve the detection performance of small-scale defects in infrared images, the network structure is optimized by introducing an attention mechanism, which suppresses global thermal diffusion information irrelevant to defects and enhances local defect-related features. The results indicate that the proposed method achieves improved detection accuracy and stable performance, demonstrating its applicability for rapid inspection of weld surface defects.