A new mineral quantification method via global feature fusion of plane-polarized and cross-polarized light images: MQM-p/xpl

The accurate and efficient identification of minerals in rock thin sections is essential for process mineralogy and ore characterization. However, traditional manual analysis is inherently subjective, time-consuming, and heavily dependent on expert interpretation. To address this challenge, a mineral quantification method based on global feature fusion is proposed, in which plane-polarized light (PPL) and multi-angle cross-polarized light (XPL) images are utilized to construct multidimensional feature representations, enhancing the discrimination of complex mineral phases. To leverage the complementary feature characteristics of PPL and XPL inputs, a two-stage augmentation strategy is introduced, consisting of general and polarization-specific enhancements to improve texture extraction in PPL and color feature perception in XPL. A multi-output joint training framework is further designed by incorporating global feature fusion and weighted inference mechanisms, enabling collaborative modeling of complementary features from PPL and XPL inputs and improving segmentation accuracy. Based on the segmented results, a pixel-level compositional quantification method is established to achieve mineral abundances estimation. The results demonstrate that the proposed model achieves an average F1-score (Ave.F1) of 80% and a mean Intersection over Union (mIoU) of 68% on test set. It is further validated on thin sections containing quartz, plagioclase, K-feldspar, mica, clinopyroxene, and hornblende, showing high consistency with manual interpretation in predicting major mineral abundances. This work contributes a novel and efficient approach for automated, objective, and reproducible mineralogical analysis, with potential applications in digital petrography, process mineralogy, and intelligent ore characterization.