Rock thin section segmentation by SAM

Conventional SAM-based segmentation methods for rock thin sections rely on manually defined point or box prompts, which limits efficiency and scalability. Fully automatic SAM segmentation, while prompt-free, often results in over-segmentation due to the heterogeneous texture and composition of rock samples. To address these limitations, this study proposes a segmentation approach that integrates SAM with automatically generated prompt boxes, which are derived from the grayscale distribution of each image. These prompts guide the SAM model to focus on dominant rock constituents, reducing irrelevant detail and improving segmentation accuracy and efficiency.