InterPore2025

A semi-supervised learning framework for multi-mineral segmentation of digital rocks utilizing a few labeled slices

20 May 2025, 12:05

15m

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

Speaker

Zhihao Xing (China University of Petroleum (East China))

Description

Accurate identification or segmentation of multiple minerals within digital rock images is critical for ensuring the reliability of subsequent analyses. Recently, deep learning models have remarkably improved the accuracy and efficiency of segmentation. However, supervised learning methods necessitate a large volume of segmented labels for training, while unsupervised learning methods lack the capability to automatically segment complex images. To overcome these limitations, we propose a novel model, DUNet, which achieves precise and automatic multi-mineral segmentation with a mere fraction of labeled samples. The model incorporates a simple yet effective semi-supervised learning paradigm to fully leverage the abundant unlabeled data and avoid overfitting. Furthermore, we conduct comprehensive experimental comparisons to evaluate the performance of various segmentation backbones constructed on Convolutional Neural Network (CNN) and Transformer architectures. Drawing on insights from these experiments, we design a Deformable Convolution layers-based backbone tailored for fine-grained segmentation. In the five-phase segmentation dataset of Bentheimer sandstone, our DUNet achieves a mean Intersection Over Union (mIoU) score of 0.901 trained on merely 0.5% of the labeled data, substantially outperforming the fully supervised UNet++ which attained a score of 0.828. Visual validation demonstrates that our model captures multi-scale features, providing more accurate segmentation details. Ultimately, we introduce an adaptive sampling strategy coupled with a dynamically weighted pixel-wise loss to mitigate the under-segmentation of minority mineral classes. The model minimizes user bias and manual intervention, aligning closely with mineral composition data obtained through Nuclear Magnetic Resonance (NMR) analysis.