The accurate characterization of rock and fluid properties in porous media of oil reservoir using thin sections depends on reliable segmentation and classification of the involved phases. However, in tight sandstone reservoirs, rock image segmentation and classification become a challenging task due to the limitation of resolution on representing the pores, throats, and other minerals at the microscale size. The resolution of an image has thus become a critical factor. Therefore, this study aims to use the super-resolution technique, which can enhance the image resolution by deep learning methods to overcome the limitation of original image resolution, and significantly improve the traditional segmentation and classification results. A Generative Adversarial Network (GAN)-based Super-resolution(SR) model was used as a pre-processing step to enhance the resolution of a given image. Then the effectiveness of super-resolution technique in post-processing procedures like segmentation and classification is evaluated using a tight sandstone data set from Ordos basin. The performance of segmentation with super-resolution enhancement is compared among Level set, Simple Linear Iterative Clustering (SLIC), and Watershed. As numerical test results reflect, segmentation in super-resolution image can achieve the segmentation of tiny minerals, pores, throats, and other blurry edges which can't be correctly segmented in the original image. Furthermore, in classification, we use a Convolutional Neural Network (CNN) model and a logistic regression model to demonstrate the advantage of SR enhancement. After the super-resolution process, the classification accuracy of CNN and logistic models have both improved for over 10%. The comparisons and analyses are presented to show that super-resolution could significantly improve these post-processing procedures. In addition, we discuss the potentials and limitations of applying super-resolution in segmentation and classification.
