Redes neurais convolucionais para classificar a maturação de frutos de café

• Ano de defesa: 2022
• Autor(a) principal: Santos, Darlisson Medeiros
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Agricultura e Informações Geoespaciais
• Palavras-chave em Português: MobileNet; MobileNetV2; NASNetMobile; Colheita do Café; Inteligência Artificial; Coffee Harvest; Artificial Intelligence; CNPQ::CIENCIAS AGRARIAS::AGRONOMIA::FITOTECNIA::MECANIZACAO AGRICOLA; Agronomia; Café - Cultivo; Inteligência artificial; Redes neurais (Computação); Café - Maturação
• Link de acesso: https://repositorio.ufu.br/handle/123456789/36614; http://doi.org/10.14393/ufu.di.2021.359

The price of coffee is strongly influenced by the quality of its fruits. This quality is built during its development, by providing adequate nutrition, climate, and health conditions. Defining the best time to harvest is fundamental to ensure this quality. The use of technology in agriculture has contributed to the rural producers for the development of management techniques that allow better growth and development of the plantations. The objective of this work was to develop a computational method using Convolutional Neural Networks, capable of identifying coffee fruits and their ripeness, which will later be embedded in a mobile application to define the ideal time to harvest coffee, helping the producer in the decision-making process. The work was developed using images of coffee fruits from rural properties in the cities of Romaria-MG and Monte Carmelo-MG. The images were collected following the physiological maturation of the coffee fruits. These images were analyzed, selected, and labeled by five specialists, forming the image bank. After the image bank was created, three neural network models were selected to be trained with the images. After adjustments, the best model was chosen to develop the prototypes. Two prototypes were developed, one classifies the image with an output of harvesting or not harvesting and the other classifies the maturation stages of the fruits. The results achieved through computer simulations were satisfactory, the harvest or not to harvest prototype reached accuracy results higher than 92%.