Abstract: In the current geotechnical engineering, manual identification methods are usually used to distinguish the types of rock samples, which are highly specialized, time-consuming,and susceptible to subjective factors, resulting in low accuracy. With the development of computer technology, machine learning is gradually applied to the automatic lithology identification, which opens up a new path for rock sample classification. Based on this, the authors took the meso-images of the four typical rock samples (mudstone, sandy mudstone, argillaceous sandstone, and sandstone) in the main urban area of Chongqing as the research object. Based on the Inception V3 convolutional network model and migration learning algorithm, a deep learning model of rock sample mesoscopic images was eatablished, and training and learning were completed. The results show that after 1 000 times training, the classification accuracy rate in the training set reaches 92.77%, and the classification accuracy rate in the verification set is 76.31%. Among them, the sandstone recognition accuracy rate is 97.28%, the mudstone recognition accuracy rate in the verification set is 81.85%, the argillaceous sandstone recognition accuracy rate is 72.59%, and the sandy mudstone recognition accuracy rate is 72.35%. Compared with the existing machine learning methods, this recognition model can automatically recognize the rock samples with very similar lithology, and has higher recognition accuracy, robustness, and generalization ability.

Key words: rock sample meso-image, deep learning, lithology identification, convolutional neural network, automatic classification