From visual diagnostics to deep learning: automatic mineral identification in polished section images

Studying mineralogical composition of ores is a fundamental step in the exploration of new deposits, as it allows determining the forms in which useful components are found, the processes of ore formation, and the potential recoverability of valuable elements. The mineral associations, textures, and structures of ores not only provide key information about the geology of a deposit, but also determine the choice of beneficiation methods. Despite the development of modern analytical tools and existing solutions for automatic mineral diagnosis, such as those based on the SEM-EDS method, optical microscopy remains the most accessible means of quantitative mineralogical analysis. However, it remains labor-intensive and requires highly skilled specialists. In addition, its visual nature limits the accuracy and reproducibility of results, creating a need for more effective approaches. One promising area is the automation of ore mineral identification based on images of polished sections. The aim of the work was to develop and validate a universal segmentation model based on deep learning. In the course of the research, related problems were also solved, including the creation of an open LumenStone dataset, the development of color adaptation methods, joint analysis of PPL and XPL images, panorama construction, and the development of a fast annotation method. The work applied convolutional neural network architectures, color correction and joint image processing algorithms, as well as an original sampling method that compensates for class imbalance. The proposed segmentation model demonstrated high accuracy (IoU up to 0.88, PA up to 0.96) for nine minerals. The obtained results confirmed the effectiveness of integrating deep learning and modern image processing algorithms in mineralogical analysis systems and laid the foundation for further development of digital methods in automated petrography.