An Efficient Pore Annotation Framework for Tight Sandstone Images with Segment Anything Model

Abstract

Analyzing pore structures in tight sandstone thin sections is pivotal for assessing reservoir quality and predicting hydrocarbon migration, providing critical guidance for the formulation of energy extraction strategies. However, the significant variability in pore morphology, scale, and quantity renders pixel-level manual annotation exceedingly challenging. To address this challenge, we have proposed an efficient pore annotation framework, EPAF. This framework leverages the zero-shot generalization capability of SAM, integrating category information into the feature space through category-mapping fine-tuning, automatically generating instance masks and corresponding category-specific feature embeddings for unannotated images. Subsequently, high-dimensional feature mapping is employed to ensure accurate association between category-specific features and instance masks, while clustering algorithms are utilized to achieve precise assignment of mask category labels. Experimental results demonstrate that EPAF significantly reduces annotation costs while achieving more comprehensive and accurate pore annotation. Our source code and datasets are publicly available at https://github.com/wudi-ldd/EPAF, hoping to inspire the community.