Optimized generative adversarial network for efficient resolution enhancement of 3D segmented rock tomography

Area of Science:

• Geoscience
• Computational Imaging
• Machine Learning

Background:

• 3D micro-Computed Tomography (micro-CT) is crucial for rock analysis but faces limitations in resolution and segmentation accuracy due to X-ray attenuation.
• High memory consumption in 3D deep learning hinders achieving high Super-Resolution (SR) in volumetric data.
• Accurate mineral differentiation and pore characterization are essential for Digital Rock Physics (DRP) simulations.

Purpose of the Study:

• To present a memory-efficient algorithm for enhancing segmented 3D micro-CT rock images using a Machine Learning (ML) Generative Model.
• To significantly increase image resolution and correct segmentation inaccuracies in micro-CT data.
• To overcome memory bottlenecks in 3D deep learning for high-resolution volumetric reconstructions.

Main Methods:

• Implemented a 3D Octree-Based Progressive Growing Deep Convolutional Wasserstein Generative Adversarial Network with Gradient Penalty (3D OB PG DC WGAN-GP).
• Utilized memory-efficient 3D Octree-Based Convolutional layers via the Minkowski Engine library to address high memory consumption.
• Trained the model using segmented 3D Low-Resolution (LR) micro-CT images and unpaired 2D High-Resolution (HR) Laser Scanning Microscope (LSM) images.

Main Results:

• Achieved a 16× increase in image resolution, improving from 7 µm/voxel to 0.44 µm/voxel.
• Successfully corrected segmentation inaccuracies caused by overlapping X-ray attenuation, enabling accurate mineral differentiation.
• Generated high-quality, segmented 3D SR images, demonstrating substantial improvements in pore characterization on Berea sandstone.

Conclusions:

• The proposed memory-efficient algorithm significantly enhances the quality and resolution of segmented 3D micro-CT rock images.
• The Octree structure effectively overcomes memory limitations in 3D deep learning, enabling unprecedented 16× Super-Resolution.
• This framework provides a robust solution for high-resolution 3D geoscientific imaging, advancing DRP simulations and mineral analysis.