Abstract: Objective Application of image processing technology in pipeline’s weld recognition system has become main application direction of machine vision in weld detection. Identification of surface defects on welds is a key technology for the application. In order to improve recognition effect of surface defects on weld, it is necessary to effectively segment weld images. In response to possible blurriness of weld boundary area inside pipeline, which leads to inaccurate segmentation results, corresponding techniques need to be adopted to improve it. Methods An improved U-Net image segmentation method was proposed to solve the problem of image segmentation of pipeline inner weld defects. Taking images of weld inside pipelines as the research object, the improved U-Net network was used to recognize and segment defect images of weld inside pipelines. After network training and model testing, segmentation results were compared with original U-Net network and FCN network. Results The results showed that the two evaluation indexes of similarity coefficient (Dice) and mean intersection over union (mIoU) of the improved U-Net network in the segmentation of weld defect images inside pipelines reached 0.8420 and 0.8514 respectively. 13.44% and 8.68% were improved respectively compared with FCN network, and 6.51% and 3.31% were increased compared with original U-Net network. Conclusion Therefore, the improved U-Net network proposed in this paper had a better effect on identification and segmentation of pipeline’s weld defects, and also provided a reliable basis for the study of pipeline’s weld defect identification system, reducing cost and time of manual detection.