Application of Machine Learning to Recognize Plagioclases in Thin Sections
Marina Ya. Kuzina · Вестник Югорского государственного университета · 2025
In this paper, YOLOv8 series models for recognizing plagioclases in thin sections are studied. Subject of research: application of algorithms for mineral recognition (in this case, plagioclases) under the microscope. Purpose of research: is to identify the best model for mineral identification and to select databases for optimal model operation. Research methods: calculation and comparison of performance metrics, analysis using neural network models of labeled images of rocks taken on a polarizing microscope with the analyzer turned on, and containing plagioclases. Object of research: algorithms for pattern recognition. Research findings: data on the effectiveness of various models were obtained, the metrics Precision, Recall, mAP50 and mAP50-95 were calculated. The best results of mineral recognition were shown by the YOLOv8n model, the accuracy of object detection was 0,808. The YOLOv8 series models were also trained on an extremely small amount of data (20 photo), metrics were obtained and their operation was tested under such conditions.