In response to the problems of traditional defect detection algorithms, such as poor accuracy and feature loss in practical applications due to the inconspicuous characteristics of welding defects and complex background information, this paper proposes a welding surface defect detection algorithm based on the improved YOLOv8 (GD-YOLO). The model first introduces the fusion of feature extraction modules and convolutional modules to enhance its information extraction capabilities. Then, a slim-neck structure is embedded in the neck network, and the upsampling operator CAFARE is referenced in the feature fusion stage to assist in enhancing the model's performance. Subsequently, the attention mechanism module is improved to optimize the overall performance without significantly increasing the computational burden. Finally, the loss function is changed to Inner-SIOU to address the problem of mismatched bounding boxes. Experimental results show that the mAP0.5 detection metric of the model in this paper is 7.8% higher than that of the baseline model, and the number of parameters and the amount of computation are reduced by 0.2 M and 0.7 G, respectively.