Classificação de imagens de soja baseada em aprendizado profundo
Brazil is one of the biggest soy producers and exporters in the world. Several tests are carried out on quality control of the seed industry, always striving to guarantee excellence in seed quality. Among the several tests used in the seeds laboratories, the tetrazolium test stands out due to its ac...
| Autor principal: | Souza Junior, Marcelo de |
|---|---|
| Formato: | Dissertação |
| Idioma: | Português |
| Publicado em: |
Universidade Tecnológica Federal do Paraná
2019
|
| Assuntos: | |
| Acesso em linha: |
http://repositorio.utfpr.edu.br/jspui/handle/1/4506 |
| Tags: |
Adicionar Tag
Sem tags, seja o primeiro a adicionar uma tag!
|
| Resumo: |
Brazil is one of the biggest soy producers and exporters in the world. Several tests are carried out on quality control of the seed industry, always striving to guarantee excellence in seed quality. Among the several tests used in the seeds laboratories, the tetrazolium test stands out due to its accuracy and speed, as well as delivering information regarding the feasibility and vigor (vitality) of seeds lots. The vitality is one of the most important characteristic of the seeds, since it determines the potential for the plants to germinate, emerge and result in normal plantules. Nevertheless, the classification of the seeds’ vigor is totally linked to the knowledge and experience of the seeds analyst. This visual analysis is a highly tiresome and, thus, an errorprone task. Due to this issue, the present project aimed to develop a framework capable provide an automatic classification of soybean seed vigor (submitted to the tetrazolium test) through its damages, as well as their severity levels, hence helping the seed analyst. To do so, the proposed frammework integrates computer vision techniques and machine learning approches. The feature extraction process is accomplished not only through hand-crafet features, but also using deep features obtained from different convolutional neural networks with transfer learning techniques. Morevoer, it was analyzed the behavior of several supervised classifiers joined with the aforecited features type regarding the seed vigor classification. The obtained results testified that it was possible to obtain good results related to the classification acuracy of the seeds’ vigor, and also important conclusions to this context. |
|---|