Utilising convolutional neural networks to perform fast automated modal mineralogy analysis for thin-section optical microscopy

Thin section microscopy has been historically used for modal mineralogy in exploration and for monitoring plant performance. Despite this, the technique relies on visual detection from expert mineralogists which is error prone and slow. Consequently, mineralogy characterisation has been largely replaced by automated mineralogy solutions like Scanning Electron Microscopy (SEM) and Energy-Dispersive X-ray Spectroscopy (EDS) which rely on chemical composition differences from electron-sample or x-ray spectrums respectively. However, these techniques are limited when minerals of interest have similar chemical compositions but different optical reflectance properties. This study aims to utilise deep learning algorithms to overcome the limitations of thin section microscopy by automating the process for quicker identification. This is done through Convolutional Neural Networks (CNNs), which are deep learning-based instance segmentation algorithms capable of detecting grain boundaries and classify minerals. Through this methodology, the proposed algorithm can evaluate grade by mineralogy compared to elemental grade in other commonly used methods. In this study, two instance segmentation algorithms were compared, namely Mask R-CNN and SOLO v2 in their ability to identify, and segment minerals to estimate surface modal mineralogy and particle size distribution and speed. The SOLO v2 algorithm achieved superior performance (APCOCO = 50.4% vs APCOCO = 44.3%) and can segment 640 × 480 thin section microscopy images at a speed of 20 frame/second which is four times faster than Mask-RCNN. This is equivalent to 120,000 grains/minute which is 2,000 times and 180 times faster than a human expert and SEM/EDS respectively.