Weld defects inside pipelines are the main cause of pipeline leakage and rupture accidents, and X-rays can effectively detect these defects. However, weld defects have problems such as multiple types, small sizes and complex backgrounds, which affect the detection accuracy. To address the current deep learning-based weld defect detection model"s insufficient adaptability to image complex background and lighting changes, and poor detection of small targets. We add the channel attention mechanism and modify the residual block structure on the backbone network of Faster R-CNN network, and adopt the improved model of ROI align to replace the ROI Pooling of traditional Faster R-CNN network. The experimental results show that compared with the original algorithm, the improved Faster R-CNN network model improves the mean Average Precision (mAP) and F1 score by15.82% and 15.82%, respectively, which can meet the requirements of high-precision detection of weld defects, and it has important theoretical significance and good It has important theoretical significance and good prospects for engineering application.