Abstract: Based on convolutional neural networks, an internal corrosion classification and recognition model was constructed by establishing the relationship between corrosion features and corrosion types. This method took macroscopic corrosion morphology images of pipelines as input, used a convolutional neural networks to extract important features from the images, and employed a classifier to categorize these features. Finally, the Grad-CAM method was applied to visualize and analyze the last convolutional layer of DenseNet-121 model. The results show that the DenseNet-121 model achieved an accuracy of 0.792. The visualization analysis revealed that the features extracted by the model were indeed key features for discriminating internal corrosion categories, and the classification basis was aligned with that of corrosion experts.