Super-resolution of real-world rock microcomputed tomography images using cycle-consistent generative adversarial...

Area of Science:

• Geoscience and Materials Science
• Image Processing and Artificial Intelligence

Background:

• Microcomputed tomography (MCT) is crucial for rock microstructure analysis.
• MCT faces a trade-off between field of view (FOV) and resolution, limiting detailed analysis.
• Existing super-resolution (SR) methods struggle with real-world rock MCT data due to reliance on paired images.

Purpose of the Study:

• To develop an effective super-resolution (SR) method for microcomputed tomography (MCT) rock images.
• To overcome the limitations of paired image requirements in traditional SR techniques.
• To enable the generation of both large field of view (FOV) and high-resolution (HR) rock MCT images.

Main Methods:

• Proposed a cycle-consistent generative adversarial network (CycleGAN)-based SR approach, named SRCycleGAN.
• Trained SRCycleGAN using unpaired low-resolution (LR) and high-resolution (HR) rock MCT images.
• Applied the trained SRCycleGAN to enhance LR rock MCT images in an online testing phase.

Main Results:

• SRCycleGAN significantly improved the quality of both simulated and real-world rock MCT images.
• Reconstructed HR images demonstrated strong agreement with target data in visual quality and statistical parameters (e.g., porosity, pore size distribution).
• The method successfully generated large FOV and HR rock MCT images, overcoming imaging system limitations.

Conclusions:

• SRCycleGAN offers a robust solution for super-resolution in rock MCT imaging, particularly for real-world datasets.
• The approach effectively models the resolution mapping using unpaired data, addressing a key challenge in the field.
• This work facilitates obtaining high-resolution, large field-of-view rock images, enhancing the study of rock properties and microstructure.