Abstract: The correct detection of surface defects of metal welds is the prerequisite to ensure the safe use of workpieces. Because the defects are similar in color to the base material and the images are not clear, it is difficult to completely detect all defect categories using conventional 2D RGB vision, and depth information needs to be added to assist detection. Therefore, a weld surface defect detection method based on RGB-D data feature fusion is proposed. Based on YOLOv8 network model, this method extracts effective feature layers of RGB and depth features by using improved symmetric backbone network structure, and introduces RGB-D data feature fusion module to realize the fusion of RGB and depth features in space and channel position. The CIOU-NMS non-maximum suppression module was added to YOLOv8 model to improve the accuracy of the check box. In this paper, the experiments were carried out in four categories: burn through, spatter, weld nog and stomata. The results showed that the missed detection rate of the improved YOLOv8 was reduced by 17.84% and the false detection rate by 19.54% compared with YOLOv8. The effectiveness and accuracy of the proposed method are proved.