Abstract: To address the engineering challenge of intelligent lithology identification based on images, this study involves photographing rock blocks at railway tunnel construction sites, conducting hand-specimen identification, and assigning lithology labels to the rock images in combination with relevant geological data. Image enhancement techniques, such as rotation and flipping, were employed to construct a lithological image dataset. Subsequently, the classification performance of three convolutional neural networks (DenseNet121, InceptionV3, and ResNet50V2) on this dataset was compared. Ensemble learning was utilized to investigate the integration effects of these three base learners under various combination strategies and with different meta-learners (Logistic Regression, Support Vector Machine, and Decision Tree). The classification predictive capability of the optimal ensemble model was then distilled into a lightweight model (MobileNetV2). By setting different distillation temperatures, the optimal student model was selected. Finally, generalization was validated using both public datasets and newly collected rock block images.