InterPore2023

Pseudo 3D unpaired domain transfer network for digital rock domain adaptation

23 May 2023, 10:00

15m

Oral Presentation (MS15) Machine Learning and Big Data in Porous Media MS15

Speaker

Mr Kunning Tang (University of New South Wales)

Description

Analyzing the physics under the same imaging condition is hampered by the domain difference between digital rock images from micro-computed tomography (micro-CT). Different scan devices, scan conditions, and sample conditions (dry/wet samples) are frequently to blame for domain differences in micro-CT rock images. Unpaired domain transfer by Generative Adversarial Network (GAN) is a method that reduces domain differences by transferring the image style from one to another without the requirement of paired images. Herein, we develop a pseudo-3D domain transfer network, Pseudo-3D Semantic CycleGAN (3D-PSCycleGAN) that transfers the rock domains with the user-defined semantic information in a 3D manner while only requiring 2D computational resources. The 3D stacking effect that is present in 2D networks without fail is eliminated by the pseudo-3D transmission. The 3D-PSCycleGAN opens up a way to analyze digital rock images under the same condition to avoid any bias or inconsistency.