Deep learning aplicado à inspeção visual de madeira, pólen e vírus

Detalhes bibliográficos
Ano de defesa: 2021
Autor(a) principal: Geus, André Reis de
Orientador(a): Não Informado pela instituição
Banca de defesa: Não Informado pela instituição
Tipo de documento: Tese
Tipo de acesso: Acesso aberto
Idioma: por
Instituição de defesa: Universidade Federal de Uberlândia
Brasil
Programa de Pós-graduação em Ciência da Computação
Programa de Pós-Graduação: Não Informado pela instituição
Departamento: Não Informado pela instituição
País: Não Informado pela instituição
Palavras-chave em Português:
Link de acesso: https://repositorio.ufu.br/handle/123456789/32470
http://doi.org/10.14393/ufu.te.2021.280
Resumo: Pollen, virus, and wood species recognition has been shown to be an important task for a number of areas, contributing to criminal investigations, climate studies, drug developments, among others. However, these studies rely on highly qualified professionals to analyze microscopic images, which have become scarce and costly. Therefore, the automation of these tasks using computational methods is promising. Recently, Deep Learning has proven to be the ultimate set of techniques for many computer vision tasks, however, it is a very difficult task to build a data set with enough samples to train these techniques from scratch. In this study, the use of transfer learning was investigated pre-training deep neural networks for pollen, wood, and virus image classification and compared their results with training from scratch and with promising pre-designed features. It also introduced the largest data sets of pollen and wood images to the present date, with 134 and 281 classes, respectively. Results indicate that even with a high number of classes, the proposed methodologies are capable of achieving acceptable classification accuracy: 96.24% in the pollen and 98.75% in the wood classification. Additionally, a new portable device combined with a new image acquisition protocol was developed to perform the wood identification outside a laboratory. Initial results evaluating 11 amazon wood species achieved 98.13% accuracy. As for the virus classification, the deep neural networks have shown to be more effective achieving 89% accuracy, being 2.8% superior when compared to the best pre-designed features study in the literature.