A benchmark dataset and baseline methods for rock microstructure interpretation in SEM images

Accurate segmentation of rock microstructures in Scanning Electron Microscope (SEM) images is essential for analyzing porosity and mineral composition. Deep learning (DL) techniques are promising to automate this segmentation, however, their effectiveness is hindered by the lack of labeled SEM datasets. To address this challenge, we developed a standardized SEM dataset including images of mudstone, sandstone, and shale. To improve the quality and diversity of the dataset, we applied preprocessing steps such as magnification standardization, median filtering, and contrast-limited adaptive histogram equalization to each SEM image. Using this dataset, we compared traditional segmentation methods with state-of-the-art DL models. Our results demonstrate that DL models significantly outperform traditional methods in capturing complex microstructural details. To facilitate further research, we publicly release the dataset along with implementations of both traditional segmentation algorithms and DL models as benchmarks, providing a valuable reference for further methodological advancements. This study not only offers a high-quality dataset but also contributes to the ongoing development of automated rock SEM image analysis through comprehensive performance evaluations.