Abstract: The end-to-end target recognition algorithm YOLO V3 was introduced into the field of weld defect detection in this paper. According to characteristics of small area and irregularity of the weld defects, K-means algorithm was used for the weld defect database to cluster and obtain the new target candidate frame and GIou, which acted two strategies of the target frame loss function to improve the original YOLO V3 network structure. Finally, a comparative test of the original YOLO V3 algorithm and the improved YOLO V3 algorithm was carried out on the weld defect data set. The loss value in the training process and the mean average accuracy in the detection process of each algorithm model were compared and analyzed. The experimental results showed that compared with the original YOLO V3 algorithm, the improved algorithm with the two strategies had a great improvement in convergence speed and better performance in pipeline defect identification, especially in average accuracy for the category of crack, incomplete fusion, incomplete penetration.