Implementação de redes convolucionais para a segmentação de imagens em tempo real com vistas à aplicação em robôs autônomos com dispositivos de visão de baixo custo

• Ano de defesa: 2018
• Autor(a) principal: Rodrigues, Carlos Alberto de Sousa Parente
• Orientador(a): Cruz Júnior, Gélson da
• Banca de defesa: Cruz Júnior, Gélson da, Vinhal, Cássio Dener Noronha, Soares, Fabrízzio Alphonsus de Melo Nunes, Silva, Karina Rocha Gomes da
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Goiás
• Programa de Pós-Graduação: Programa de Pós-graduação em Engenharia Elétrica e da Computação (EMC)
• Departamento: Escola de Engenharia Elétrica, Mecânica e de Computação - EMC (RG)
• País: Brasil
• Palavras-chave em Português: Redes neurais artificiais; Visão computacional; Segmentação de imagens
• Palavras-chave em Inglês: Artificial neural networks; Computer vision; Image segmentation
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.bc.ufg.br/tede/handle/tede/8811

This work presents a study of convolutional networks to segment and classify images. The purpose of this network is to eventually give more autonomy to LEIA 1 robot, using the computer vision information in its processing. Methods such as this attempts to adapt the visual perception system of living beings. The complexity of this task lies in not having sufficient understanding of the biological system to model a system capable of processing images with the same speed and efficiency as a human. To accomplish this work, two different convolutional network architectures were validated. The first network has 13 layers, while the second has 15 layers, and more adjustable weights than the first one. For training and validation, a slice of Playing for Data dataset was used and adapted. The training set was composed of 300 images, and the network was validated using 2500 patterns. For each architecture, three training routines were performed, using the Adam, Nadam and Adamax methods. The most relevant results used the 15-layer architecture with Adamax optimizer.