Case study of deep learning image segmentation for the purposes of rapid 2D petrographic analysis in volcanic rocks

Automation using deep learning methods is a useful alternative to manual methods of petrographic segmentation, but often requires user familiarity with coding and/or algorithms. We examine the Dragonfly^TM program's deep learning tools for application by users with a variety of skill levels as a method for petrographic image segmentation. An image processing methodology, bimodal image stacking, was created for low-input-data, high-efficacy training of models which can then be applied to varied samples. Using backscatter electron images we show that the resulting model segmentations agree with manual segmentation total and modal crystallinity values within 5%, and calculated plagioclase crystal size distribution (CSD) values within 2σ, despite limitations in discriminating mafic phases. Model creation and training takes