deployment of the model and identification-measurement functions through an intuitive graphical interface. The input data consist of microphotographs of thin sections of intrusive igneous rocks in JPG, JPEG, or PNG format, in the RGB color system, along with a CSV file containing the conversion factors for each image. This file may have any name and must include three columns: Image, containing the image name and file extension; Conversion_mm, with the linear conversion factor (mm/px); and Conversion_mm2, with the areal conversion factor (mm²/px). The output consists of an automated report including mineralogical analysis, a ternary diagram, a segmented mineral image, and histograms of major-axis diameter and area distributions. An individual report is generated for each image. Figure 9 – AI Mineral Tech web app. (a) Google Colab version, (b) Hugging Face, (c) Automatic report. Available at: https://stalynpaucar271828.wixsite.com/auto-descripcion 6. CONCLUSIONS AND IMPLICATIONS FOR INDUSTRY A model that combines the Detectron2-v0.6 architecture with the scikit-image library has been implemented to identify and measure minerals on thin sections of intrusive igneous rocks. For the training and validation of Detectron2, a dataset of 400 JPG images in RGB color, both in plane-polarized light (PPL) and cross-polarized light (XPL), was generated from fresh or slightly altered intrusive igneous rocks within the QAP diagram of Streckeisen (1976). The labeling set is heterogeneous, with plagioclase (Pl) being the most abundant with 21150 labels, and olivine the least frequent with 728 labels. Among the training results, the XPL-10X model stood out for its superior performance, achieving a total loss of 1.15. For validation, the average precision (mask AP) of 39.5% from the ResNetX101 network was used as a reference. The XPL- (a) (b) (c)
RkJQdWJsaXNoZXIy MTM0Mzk2