Abstract: Aiming at the problems of low detection accuracy and slow detection speed caused by small target and complex background of steel pipe welding defects, an improved YOLOv5 detection algorithm is proposed. First, recursive gated convolution gnConv is used to replace the common convolution layer in the network, which enhances the interaction ability of model space, realizes efficient feature extraction, and indirectly improves the detection speed. Secondly, the use of ASPP (Atmosphere Spatial Pyramid Pooling) module not only expands the receptive field, but also improves the detection speed. Finally, GAM (Global Attention Mechanism) is added to the prediction end of the network to further enhance feature extraction and improve detection accuracy. The experimental results show that the improved al gorithm mAP achieves 92.7%, 2.1 percentage points higher than the original algorithm, and the speed is 50.8 f/s, meeting the requirements of precision and real-time of steel pipe welding defect detection.