Automatic classification method of rock lithology based on ResNet network and deep transfer learning

[Objective]To explore the applicability of deep learning frameworks in lithology recognition and address the limitations of traditional methods characterized by low efficiency and strong subjectivity, an automatic classification method for rock lithology based on a ResNet convolutional neural network combined with transfer learning is proposed. Six types of rock images, including granite, marble, quartzite, limestone, coal rock, and sandstone, are selected for experimental analysis. [Methods]A dataset containing 7, 416 rock images is constructed through data augmentation and divided into training, validation, and test sets. In model development, ImageNet pre-trained weights are introduced, and multiple transfer learning strategies are designed. Comparative experiments are conducted on ResNet-18, ResNet-34, and ResNet-50 models. Meanwhile, batch normalization, learning rate decay, and the Adam optimizer are employed to improve network performance. [Results]The results indicate that under small-sample conditions, the fully fine-tuned ResNet-18 model achieves the best performance, with an accuracy of 96.10%, precision of 96.01%, and recall of 96.12%, outperforming the other models. [Conclusion]Compared with the other two models, the proposed model demonstrates higher classification accuracy, faster convergence speed, and stronger robustness in recognizing complex lithological features. It significantly improves training efficiency and successfully realizes automatic lithology classification, providing an effective technical support for geological exploration and engineering applications.