Paired and Unpaired Deep Learning Methods for Physically Accurate Super-Resolution Carbonate Rock Images

Abstract X-ray micro-computed tomography (micro-CT) has been widely leveraged to characterise the pore-scale geometry of subsurface porous rocks. Recent developments in super-resolution (SR) methods using deep learning allow for digital enhancement low-resolution (LR) images over large spatial scales, creating SR comparable high-resolution (HR) ground truth images. This circumvents common trade-off between resolution and field-of-view. An outstanding issue is use paired LR HR data, which often required training step such but difficult obtain. In this work, we rigorously compare two state-of-the-art techniques, both unpaired with like-for-like data. The first approach requires train a convolutional neural network (CNN), while second uses generative adversarial (GAN). approaches are compared micro-CT carbonate rock sample complicated micro-porous textures. We implemented various image-based numerical verifications experimental validation quantitatively evaluate physical accuracy sensitivities methods. Our quantitative results show that GAN can reconstruct as precise CNN method, times dataset requirements. unlocks new applications image methods; registration no longer needed during data processing stage. Decoupled from storage platforms be exploited networks applications. opens up pathway related multi-scale flow simulations heterogeneous media.