Abstract: Aimed at the problems of large labor consumption and low efficiency in metal surface defect recognition in industrial production, an improved VGG16 network metal surface defect image classification method is proposed. Based on the VGG16 network, the attention mechanism CBAM was introduced to enhance the feature learning ability, and the Inception network structure was introduced to broaden the network width, thereby enhancing the nonlinear ability of the model. The input image data was processed to improve the robustness of the network model. Through experimental verification, the improved network model has an accuracy rate of 90.23% on the data set GC10-DET, and an accuracy rate of 98.84% on the data set NEU-CLS. The experimental results show that this method has good practical application significance in the classification of metal surface defects.