Enhancing digital rock image resolution with a GAN constrained by prior and perceptual information

Abstract

In digital rock physics, pore distribution and microstructure of rocks affect the flow properties and elastic characteristic of rocks. It is difficult to accurately characterize these parameters in low-resolution images due to rock heterogeneity, and improving with instrument is costly. Traditional interpolation method to enhance the image resolution only enriches the low frequency information of the image without focusing on the boundary which is rich in high frequency information, and conventional deep-learning methods cannot perform a satisfying result in different-level instrument noise. To address these issues, a generative adversarial network of an image segmentation network as discriminator constrained by perspective information and prior information (SCPGAN) is proposed to improve images resolution by enhancing medium-high frequency information of images and improving network anti-noise capacity. The perceptual information is extracted by VGG19 to enrich boundary and texture information and the porosity is built by discriminator as prior information to guide generator to improve noise immunity. Traditional interpolation and CNN are used to verify that perceptual information can be augmented with medium-high frequency information, and GAN-based models are used to verify that prior information can improve network performance. The result shows the network with perceptual information can significantly improve images resolution and outperform the traditional interpolation on structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and multiple-point connectivity (MPC), while GAN-based model with prior information shows an excellent anti-noise capacity, especially high-level noise, and has similar physical properties to the high-resolution model in the permeability simulation.