Metal mineral classification under microscope images using deep learning

Analysis of optical microscopic image data is crucial for identifications of mineral phases, and thus directly relevant to the subsequent methodology selections of the further detailed petrological exploration. So far, large-dimensional image analyses are dominantly based on digital image datasets, and the automatic identification of the optical microscopic data is still poorly examined. Here, by testing the Swin Transformer, a deep learning algorithm on different metal mineral phases, we proposed a well-behaved mineral recognizer with high accuracy of 92.8% and strong global ability. In addition, we apply Class Activation Mapping (CAM) is introduced for the first time in mineral identification tasks and reveals the interpretability of the models, allowing us to more intuitively observe that mineral edges are the most effective model identification features. The results demonstrated that optical microscope data can not only rely on pixel information, and machine learning can accurately extract all available attributes, which reveals the potential to assist in data exploration and provides an opportunity to carry out spatial quantization at a large scale (cm-mm).

Keywords: Metal mineral; Microscope images; Deep learning; Swin Transformer; Class Activation Mapping