Deep Learning Domain Adaptation Applied for Minerals Semantic Segmentation in Reflected Light Microscopy Images

In the mining industry, mineral characterization provides data and parameters to support efficient and profitable ore processing. However, mineral characterization techniques usually require extensive image analysis, making manual large-scale image segmentation of mineral phases impractical. Considering the accuracy level currently achieved with deep learning models, they represent a potential solution to the problem of automating mineralogical ore characterization. However, training deep learning models generally requires an abundance of annotated images. Additionally, supervised learning models trained on data of a given ore sample tend to perform poorly on a sample with different characteristics, or of a different ore. In this work, we consider those different samples as pertaining to different domains: a source domain, used for training the model, and a target domain, in which the model will be tested. In such application context, domain divergences, also regarded as domain shift, may emerge from differences in mineral composition, or from distinct sample preparation processes. This research evaluates the use of the unsupervised deep domain adaptation to obtain models that generalize properly for a target domain even though no labeled target domain samples are used during training. The task of the models is to discriminate between ore and resin pixels in reflected light microscopy images. Preliminary cross-validation experiments between different domains prior to domain adaptation revealed a pronounced difficulty in the models' generalization. This fact motivates the herein presented research regarding evaluation of the potential of domain adaptation as an attempt to compensate for the loss of performance caused by domain shifts. The results of the domain adaptation showed that a significative part of the adapted models presented performance metrics considerably above the cross-validation baseline, achieving F1 score gains of up to 33% and 38% in the best cases, although in some source-target combinations limited performance gains were obtained. This indicates that the intensity of the displacement between the source and target domains may limit the success of the domain adaptation method.