Aprendizagem profunda aplicada à caracterização de minérios: discriminando minerais opacos e não opacos de resina epóxi em imagens de microscopia ótica de luz refletida

• Autor(a) principal: Filippo, Michel Pedro
• Data de Publicação: 2020
• Outros Autores: [e-mail]
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UERJ
• Texto Completo: http://www.bdtd.uerj.br/handle/1/20282
• topic: Deep learning; Image analysis; Semantic segmentation; Ore microscopy; Iron ore; Processamento de imagens; Análise de imagem; Microscopia ótica; Minérios; Aprendizagem profunda; Análise de imagem; Segmentação semântica; Microscopia de minério; Minério de ferro; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO::PROCESSAMENTO GRAFICO (GRAPHICS)

Discrimination between non-opaque minerals and embedding resin in reflected light microscopy images from ore samples is a challenging and well-documented problem. The similar specular reflectance of these materials makes it difficult to discriminate, even by human experts. Although slight visual differences, such as internal reflections and subtle differences in the polished surface, can help humans to delineate distinct particles of these materials, conventional image processing techniques are unable to capture such subjective features and tend to fail, becoming a problem for which a robust computational solution was still missing. Inspired by the recent success of deep learning techniques in image interpretation, the present work evaluates the effectiveness of semantic segmentation of ore images through deep learning models, in the discrimination between embedding epoxy resin and ore particles containing opaque and nonopaque minerals. In this work, the performance of the DeepLabv3+ architecture is evaluated and some variants are proposed in order to improve segmentation accuracy, particularly at the boundaries of mineral particles. The deep learning models were evaluated using four different data sets, containing images of different ores, acquired with different experimental configurations. The results showed outstanding 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 deep learning solution, cross-validation evaluations were conducted, using one out of the four datasets for training the model, and testing it on the other datasets. Possibly, this work presents the first approach of semantic segmentation based on deep learning for the discrimination of opaque and non-opaque epoxy resin minerals in reflected light microscopy images