Abstract: Using deep learning technology to carry out image recognition of tight sandstone reservoir thin section can effectively improve the accuracy and efficiency of compositional identification, which is the development trend of mineral identification. However, factors such as the difficulty of thin section image acquisition, the high cost of annotation and privacy protection, result in a scarcity of tight sandstone thin section image samples, which cannot meet the training requirements of deep learning image recognition models. In order to increase the number of samples and improve the training effect of the deep learning model, this paper takes the Linxing block in Ordos basin as the target area and proposes a tight sandstone image augmentation method by combining the respective advantages of oversampling data method and data deformation method. Firstly, the style generative adversarial network is improved to generate high-resolution tight sandstone images and enhance data diversity. Secondly, we use data deformation methods to achieve data augmentation with labeled images, thereby reducing annotation costs and expanding the data scale. Finally, the Blend Mask algorithm is trained using augmented data to accurately identify and precisely segment sandstone particles. The experimental results show that compared with the similar contrast algorithms, this proposed method has an IS (inception score) value with a maximum of 2.43 and an FID (Fréchet inception distance) value with a minimum of 22.71. Meanwhile, the recognition accuracy after adding the generated images is 92.7%. It indicates that the method proposed in this paper has significant advantages in terms of the quality of generated images and the improvement of the training effect of the deep learning model.

Key words: tight sandstone reservoir, thin section image, deep learning, image generation and recognition, Ordos basin