Intelligent segmentation and parameter characterization of coal particles and minerals based on SAM

Accurate extraction of microstructural information from coal is of great significance for the clean utilization of coal. However, the complex intergrowth structure of coal increases the difficulty of identifying and segmenting different components. To address the strong dependence of conventional segmentation methods on high-quality annotated data and their limited adaptability across samples, the segment anything model (SAM), based on large-scale pretraining and a prompt-driven mechanism, is introduced for coal microscopic images and adapted to the target domain. By incorporating Frequency-tuned (FT) point-prompt optimization, single-class-guided fine-tuning, and boundary loss constraint, an optimized SAM-based intelligent segmentation model for coal particles and minerals is developed. Specifically, the FT point-prompt optimization enhances the effectiveness of prompt points by using saliency maps, single-class-guided fine-tuning enables SAM to transition from class-agnostic segmentation to segmentation with mineral semantic recognition capability, and boundary loss constraint strengthens the model’s ability to learn boundary features. The optimized SAM is compared with K-means, U-Net, and DeepLabv3 models to evaluate their segmentation performance and model complexity. The accuracy of the four models in extracting four types of parameters, including coal particle size, particle shape, mineral proportion, and mineral-organic matter interface length, is compared. The cross-sample segmentation performance of the optimized SAM and its effectiveness in liberation analysis were evaluated. The results show that the optimized SAM achieves the best performance in coal particle and mineral segmentation tasks, with pixel accuracy (A) and mean intersection over union (mIoU) values of 96.80% and 92.42%, but its high-precision segmentation capability is accompanied by relatively high model complexity. Meanwhile, the application parameters extracted from the segmentation results of this model show high consistency with the reference values, with all coefficients of determination R² exceeding 0.8, while maintaining stable segmentation performance under different sample preparation conditions. Liberation statistical analysis based on the segmentation results demonstrates that the model can effectively support coal liberation evaluation, with the area-based liberation degree showing high consistency with that obtained by the standard point-counting method. Overall, the optimized SAM enables automated segmentation of coal microscopic images and effective characterization of application parameters, providing reliable technical support for coal particle structure analysis, mineral occurrence characterization, and liberation performance evaluation.