Enhancing the Resolution of Micro-CT Images of Rock Samples via Unsupervised Machine Learning based on a Diffusion Model

Objectives/Scope X-ray Micro-Computer Tomography (μ-CT) has been widely adopted in earth science and petroleum engineering due to its non-destructive characteristic. Meanwhile, this three-dimensional-imaging method can be integrated with computer simulation to investigate petrophysical properties of reservoir rocks at pore scales. However, the application of μ-CT is limited by the trade-off between field of view and resolution, and it is challenging to indicate the pore structure of rocks, especially for shale or carbonate rocks. To address this issue, deep-learning-based super-resolution techniques have rapidly developed in the past few years. Methodology In this study, a super-resolution algorithm based on the state-of-the-art (STOA) diffusion model is proposed to generate super-resolved CT images for carbonate rocks. The proposed method adapts denoising diffusion probabilistic models to conditional image generation and performs super-resolution through a stochastic denoising process. Cascaded diffusion model is utilized to increase the training speed and generate high fidelity CT images. This method exhibits superior performance in the resolution-enhancement of CT images at various magnification factors (with a large scaling factor of up to 16) without the occurrence of image-noise and image-blurring issue, and the super-resolved CT images performs well for the calculation of petrophysical properties of carbonate rocks. Results This algorithm is applied to the carbonate rock and the performance of the diffusion model is evaluated by quantitative extraction and qualitative visualization. In addition, this method is compared with other methods, such as GAN, Variational Autoencoder, and Super-Resolution Convolutional Neural Networks (SRCNN). The results indicate that the built model shows excellent potential in enhancing the resolution of heterogeneous carbonate rocks. To be specific, the super-resolved images exhibit clear and sharp edges and a detailed pore network. In addition, it performs well on different upscaling factors (up to 16) and is superior to the existing super-resolution approaches (for both supervised and unsupervised algorithms). This study provides a novel deep-learning-based method using a diffusion model to enhance the resolution of μ-CT images of carbonate rocks (up to 16). Novelty The novelty of this study is three-fold. First, this method belongs to unsupervised learning, indicating that pairs of high-resolution and low-resolution CT images are no longer needed. Second, a large scaling factor (up to 16) is reached without an image-blurring issue, which normally occurs in other deep-learning-based super-resolution algorithms. Third, the quality of super-resolved images is promising and faithful when compared with other generated learning methods, such as Generative Adversarial Networks (GAN).