Abstract: To improve the accuracy and efficiency of weld defect detection in oil and natural gas pipelines, deep learning technology was used to automatically detect defects in radiographic images of pipeline welds. The research involved preprocessing and block-wise cutting of weld images, and defect identification was performed using an improved YOLOv7 model. Model improvements included the introduction of coordinate convolution, a multi-scale feature fusion module, and an attention mechanism. Results showed that the improved model increased mAP@50 by 3.37%, significantly reduced false positives and false negatives, and enhanced the level of industrial automation and intelligence.