Adaptação de Domínio em Rede Neural Profunda para Segmentação Semântica de Minérios em Imagens de Microscopia Ótica de Luz Refletida

Abstract

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. Thus, considering the accuracy level provided by deep learning models, they represent a potential solution to the problem of automating mineralogical ore characterization. However, training deep learning models requires abundant annotated images. Besides, supervised deep learning models trained on data of a given ore sample, herein called the source domain, tend to perform poorly on a sample of another domain. In this context, the distinct domain is usually referred to as target domain, in opposition to source domain. In addition, in such application context, domain divergences, the so-called domain shift, may emerge from differences in mineral composition or even result from samples of the same mineral deposit submitted to distinct preparation processes. This research proposes the investigation of the domain adaptation technique DANN (Domain Adversarial Neural Network), based on adversarial training, to improve the classification accuracy of a DeepLabv3+ neural network model for semantic segmentation in the target domain. The task for the model is to discriminate ore and resin pixels in reflected light optical microscopy images of the target domain. Preliminary cross-validation experiments between different domains, training with data from one domain and testing on data from a different domain, revealed a pronounced difficulty in model’s 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 cross-domain shifts. The results of the domain adaptation showed that a significant part of the models trained with the DANN framework presented performance metrics considerably above the cross-validation baseline, achieving F1 score gains of approximately 38% and 33% in the best cases. However, in these same experiment set, despite that relative success, domain adaptation did not work consistently for all source and target combinations. This result may be an indication that the intensity of the displacement between the source and target domains may limit the success of the domain adaptation method.