Geo-SegNet: A contrastive learning enhanced U-net for geomaterial segmentation

Abstract

X-ray micro-computed tomography scanning and tomographic image processing is a robust method to quantify various features in geomaterials. The accuracy of the segmented results can be affected by factors including scan resolution, scanning artifacts, and human bias. To overcome these limitations, deep learning techniques are being explored to address these challenges. In the present study, a novel deep learning model called Geo-SegNet was developed to enhance segmentation accuracy over traditional U-Net models. Geo-SegNet employs contrastive learning for feature extraction by integrating this extractor as the encoder in a U-Net architecture. The model is tested using 10 feet of sandstone cores containing significant changes in porosity and pore geometries and the segmentation results are compared to common segmentation methods and U-Net. Compared to a U-Net-only model, Geo-SegNet demonstrates a 2.0 % increase in segmentation accuracy, indicating the potential of the model to improve the segmentation porosity which can also improve subsequent metrics such as permeability.