Deep learning semantic segmentation of opaque and non-opaque minerals from epoxy resin in reflected light microscopy images

Abstract

Inspired by the recent success of deep learning techniques in image interpretation, this work presents a pixel-wise semantic segmentation method for the discrimination of opaque and non-opaque minerals from epoxy resin in reflected light microscopy images, a challenging and classical problem of ore characterization. It evaluates a particular Fully Convolutional Network architecture, the so-called DeepLabv3+, and proposes extensions devised for improving segmentation accuracy, particularly at the borders of ore particles. Furthermore, this work introduces the use of correlative microscopy to objectively generate reference images (for training, validation, and test of the deep learning models) from an independent data source, instead of producing reference images through visual inspection and interactive labeling.The deep learning models were trained, validated, and tested using four distinct datasets, containing images of different ores (copper and iron ores), acquired with different experimental setups. The results showed remarkable performances, systematically over 90% overall accuracy and F1 scores, and up to 94% for some datasets. Additionally, in order to analyze the generalization capacity of the method, cross-validation evaluations were conducted.The developed method allows the automatic discrimination of ore particles (composed by opaque and/or non-opaque minerals) from epoxy resin, opening new possibilities for image analysis systems based on reflected light microscopy.