Title:
Enhanced Feature Segmentation of X-ray Micro-CT Scans of Geomaterials using Contrastive Learning and UNet-based Architecture
Author(s):
Tian
Publication:
Web Session
Volume:
ws_S24_Tian_2.pdf
Issue:
Appears on pages(s):
Keywords:
DOI:
Date:
3/23/2024
Abstract:
X-ray micro-computed (micro-CT) scanning is a robust method to quantify various features in porous geomaterials (i.e., portland cement-based materials and geological materials). However, factors like scan resolution, human bias, and drastic porosity changes along a core can impact segmentation accuracy. To overcome these challenges, deep learning techniques are being explored. In the ongoing study, a novel deep learning model designed to enhance segmentation accuracy over traditional UNet models is assessed on a sandstone core with an abrupt porosity change. The model utilizes contrastive learning for feature extraction, later integrating this extractor as the encoder in a UNet architecture. The features are then used by UNet for segmentation. Compared to a UNet-only model, this approach shows a 2 % increase in segmentation accuracy, indicating the potential to improve not only the segmentation of porosity but also other features such as high-density materials and changes in water saturation. A range of X-ray micro-CT datasets of geomaterials are being tested as ongoing work.