<p>The identification and classification of rock thin sections is an indispensable basic work in many fields, such as geological survey, engineering exploration and mineral exploration, and provides core basic support for all kinds of related research and practice. The existing rock thin section identification and classification methods based on machine learning have made important progress. However, in practical engineering applications, it is difficult to obtain a sufficient number of rock thin section images. In this paper, a novel data augmentation framework CG-AE-VAE for rock thin section identification and classification with small samples is proposed, which solves the problem of intra-class diversity of rock thin section images and alleviates the problem of insufficient training samples. CG-AE-VAE framework has excellent compatibility and adaptability, and can be embedded into the current mainstream deep learning classification model architecture. The experimental results show that in the small sample scenario where only 0.9% of the data set is used for model training, after integrating the framework into several classical machine learning models, all the integrated models show high classification accuracy, and the highest classification accuracy is 87.9%. The CG-AE-VAE framework provides a new technical path and solution for the identification and classification of rock thin section in small sample scenarios.</p>