Desenvolvimento de uma interface gráfica para utilização de algoritmos de aprendizado de máquina na classificação de imagens : um estudo de caso na seleção de grãos de sementes

• Ano de defesa: 2025
• Autor(a) principal: Magalhães Júnior, Antonio Mendes
• Orientador(a): Sáfadi, Thelma
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Lavras; Instituto de Ciências Exatas e Tecnológicas (ICET)
• Programa de Pós-Graduação: Programa de Pós-Graduação em Estatística e Experimentação Agropecuária
• País: brasil
• Palavras-chave em Português: Visão computacional; Reconhecimento de padrões; Descritores de imagens; Computer vision; Pattern recognition; Image descriptors
• Área do conhecimento CNPq: Ciências Exatas e Tecnológicas
• Link de acesso: https://repositorio.ufla.br/handle/1/60238

Brazil has consolidated its position as one of the largest global seed markets, generating bil- lions of reais each year. In this context, seed quality evaluation has become a crucial step for identifying damages — whether mechanical, physiological, or pathogenic — and for estimating their germination potential. However, the traditionally used methods, conducted by human eva- luators, present significant challenges, such as the subjectivity inherent in visual judgment and scalability limitations. Aiming to overcome these obstacles, Machine Learning methods such as Logistic Regression, Decision Trees, Random Forest, Artificial Neural Networks (ANN), and XGBoost have been applied, providing greater objectivity, speeding up the process, and redu- cing operational costs. However, the adoption of these techniques requires a considerable level of technical knowledge, which can hinder their practical implementation. This work aimed to develop an interactive graphical interface to facilitate the use of Machine Learning models for image classification, applicable not only to seeds but to any type of image with identifiable pat- terns through image descriptors. The application was built as a web tool based on the Shiny framework and the R language, allowing users with different levels of technical knowledge to complete the entire cycle of training and validation of models. The tool performs image pro- cessing techniques such as histogram equalization and segmentation and allows the selection of Machine Learning models and image descriptors to be extracted — including statistical, shape, geometric moments, and Haralick texture. Furthermore, the application enables the storage of generated models, making their future application in new analyses possible.