Processing of micro-CT images of granodiorite rock samples using convolutional neural networks (CNN), Part I: Super-resolution enhancement using a 3D CNN

September 18, 2022

A. Roslin (1), M. Marsh (2), N. Piché (3), B. Provencher (3), T.R. Mitchell (1), I.A. Onederra (1), C.R. Leonardi (1)
Minerals Engineering. Volume 188 (18 September 2022). DOI:


Convolutional neural network, micro-CT, super-resolution, igneous rocks, deep learning, U-net 3D


X-ray micro-computed tomography (micro-CT) is widely used for three-dimensional analysis of many rock types. However, the practical implementation of this method for micro-porous samples requires a compromise between the resolution of the images and the obtainable field of view (FOV). Generally, resolution enhancement results in a reduction of the FOV. The generation of high-quality micro-CT images is an expensive and time consuming task due to the competing requirements of a large FOV and fine resolution. To alleviate this, super-resolution processing, based on deep learning, is proposed to improve the quality of low-resolution images that can obtain a large FOV. In this research, a super-resolution technique employing the three-dimensional U-Net convolutional neural network (CNN) architecture was applied to enhance the resolution of granodiorite rock sample images. This was undertaken using two sets of micro-CT image triplexes, where the first triplex contained 3-, 6-, and 12-micron resolution sets, and the second triplex contained 1-, 2-, and 4-micron resolution sets. For each triplex, 80% of the images were used for training the neural network with the remaining 20% used for validation. Further validation was performed by comparing the processed results to images obtained from scanning electron microscopy (SEM). It was observed that super-resolution processing can significantly improve the low-resolution micro-CT image quality without physically reducing the sample size typically required for high-resolution scanning. It is expected that this technique could assist practitioners reveal features absent in small samples (e.g. large fractures and or rock textures). Furthermore, images restored through super-resolution processing maintain the FOV of the lower resolution scan, a task that would be comparatively expensive and time consuming to acquire in a high-resolution scan. The workflow proposed in this study has a significant impact on a range of fields including the numerical prediction of rock permeability, and segmentation for advanced mineral analysis.

How Our Software Was Used

Dragonfly’s Deep Learning Tool was used to demonstrate that a U-Net architecture can successfully be used for super-resolution.

Author Affiliation

(1) School of Mechanical and Mining Engineering, The University of Queensland, St Lucia, Australia
(2) Object Research Systems, Denver, CO, United States
(3) Object Research Systems, Montreal, Quebec, Canada