Abstract: The complexity and multiple solutions of rock thin section images lead to the difficulty in classification of rock thin sections. This paper attempts to apply the deep learning method to the classification of rock thin images. Thin section images of 6 common rock types, such as andesite, dolomite and granite, were selected in the experiment, and 1000 images of each type were used as experimental data. The VGG model was established, and the identification accuracy of the verification set reached 82% after 90, 000 iterations. Based on the analysis of the experimental data, the authors found that the rock images with similar compositions are easy to be confused; for example, dolomite and oolitic limestone are both carbonate rocks and it is easy to misjudge each other. Plagioclase porphyry, microcrystalline and cryptocrystalline or vitreous matrix were extracted from the andesite characteristic diagram, and oolitic and interstitial materials were extracted from the oolitic limestone characteristic diagram. The result obtained by the authors proves that the VGG model is effective in the classification of rock thin section.