10X model also achieved the best results in average precision AP 0.5-0.95, with 15.7%. Considering object size (MS COCO scale), all models performed well on large objects (APl = 19.9-35%) but poorly on small objects (APs = 1.7-4.0%). Initial visual evaluation of the inferences indicates moderate segmentation of the components, with gabbro showing the best results. To evaluate the accuracy of scikit-image measurements, labels from XPL and PPL images were compared, yielding 88% exact measurements and 12% with a relative error of 0-3%, confirming its suitability for measuring new labels generated by Detectron2. This library was also used to classify the size of mineral labels in the full dataset. Using the Castro (2015) scale, finegrained minerals (0.1-1 mm, 53326 labels) predominated, while according to the MS COCO scale, small (32122) and medium (25746) objects were most common. These characteristics allow for organizing a dataset that is well-structured in terms of origin, quantity, quality, format, and homogeneity, providing a valuable resource for future AI models. The development of an Artificial Intelligence model for automatic mineral identification and segmentation is a complex process, spanning from data acquisition to inference evaluation, and requiring significant time and resource investment. This study implemented a web application using Hugging Face and Streamlit, offering an accessible tool to external users. Although the current results still require expert review and are not fully autonomous, they represent a significant advancement: with more robust architectures, these models could optimize rock characterization, accelerate analyses in mining and petroleum industries, enhance geological studies, and reduce operational costs, demonstrating the potential of AI to support both academic research and industrial applications. REFERENCES Azzam, F., Blaise, T., & Brigaud, B. (2024). Automated petrographic image analysis by supervised and unsupervised machine learning methods. Sedimentologika, 2(2). Baykan, N. A., & Yılmaz, N. (2010). Mineral identification using color spaces and artificial neural networks. Computers & Geosciences, 36(1), 91-97. Bhatnagar, D. (2023). History of CNN & its impact in the field of Artificial Intelligence. Retrieved 03 21, 2024, from Medium: https://medium.com British Geological Survey. (2025). BRITROCKS: mineralogy and petrology collections database. Retrieved 02 20, 2025, from www.bgs.ac.uk/technologies/databases/bgs-rock-collections Castro, A. (2015). Petrografía de rocas ígneas y metamórficas. Paraninfo. Ehrlich, M., Davis, L., Lim, S.-N., & Shrivastava, A. (2021). Analyzing and mitigating jpeg compression defects in deep learning. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2357-2367. Girshick, R. (2018). Facebook open sources Detectron. Retrieved 04 08, 2024, from Meta: https://research.facebook.com/blog/2018/1/facebook-open-sources-detectron Guisiano, J. E., Moulines, É., Lauvaux, T., & Sublime, J. (2023). Oil and Gas Automatic Infrastructure Mapping: Leveraging High-Resolution Satellite Imagery through finetuning of object detection models. International Conference on Neural Information Processing, 442-458. Honda, H. (2020). Digging into Detectron 2 — part 1. Retrieved 04 02, 2024, from Medium: https://medium.com/@hirotoschwert/digging-into-detectron-2-47b2e794fabd
RkJQdWJsaXNoZXIy MTM0Mzk2