Abstract: The accurate detection of weld defects on the metal surface is the premise to ensure the safe use of the workpiece. Due to the similar color of the defects and the unclear image of the base metal, it is difficult to completely detect all the defect categories by using the conventional 2DRGB vision, so it is necessary to add depth information to assist the detection. This study proposes a weld surface defect detection method based on RGB-D data feature fusion. Based on the YOLOv8 network model, this method uses the improved symmetrical backbone network structure to extract the effective feature layer of RGB and depth features, and introduces the RGB-D data feature fusion module. The fusion of RGB features and depth features in space and channel position is realized. The CIoU-NMS(Complete Intersection over Union-Non Max Suppression) non-maximum suppression module is added to the YOLOv8 model to improve the accuracy of the check box. The results show that the missed detection rate of the improved YOLOv8 is reduced by 17.84% and the false detection rate is reduced by 19.54% compared with YOLOv8, which proves the effectiveness and accuracy of the proposed method.