EXTRAÇÃO DE CARACTERÍSTICAS DE IMAGENS APLICADA À DETECÇÃO DE GRÃOS ARDIDOS DE MILHO

• Ano de defesa: 2015
• Autor(a) principal: Ribeiro, Sergio Silva
• Orientador(a): Falate, Rosane
• Banca de defesa: Guimarães, Alaine Margarete , Miazaki, Mauro
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: UNIVERSIDADE ESTADUAL DE PONTA GROSSA
• Programa de Pós-Graduação: Programa de Pós Graduação Computação Aplicada
• Departamento: Computação para Tecnologias em Agricultura
• País: BR
• Palavras-chave em Português: processamento digital de imagens; mineração de dados; milho
• Palavras-chave em Inglês: digital image processing; data mining; maize
• Área do conhecimento CNPq: CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://tede2.uepg.br/jspui/handle/prefix/125

Maize is an important crop in national and international level. Its market value is influenced by the quality of the grains. The rot (damaged) grains are an economic and health problem. Strict laws determine the process of classification of grains in healthy and rot, however this procedure is done manually, subject to subjectivity and human errors. The objective of this study was to use computational methods of digital image processing for feature extraction of maize grains. This work also proposed to identify which methods and data mining algorithms are more efficient to solve the problem of rot grain maize identification. For this research it was used corn grain samples from various cooperatives in the Midwest Parana area. These samples were classified by certified technicians in these cooperatives and submitted, with manufacturer and developed programs, to the process of image acquisition, segmentation and extraction of colour and texture characteristics. For the development of programs was used Python together with SimpleCV framework. The extracted image data were saved in a file format of the Weka tool that was used to train and to test the base by employing the methods holdout and cross-validation. All tool algorithms were used for data processing and 24 of them have reached a rate equal accuracy and / or greater than 99%. The best studied related work achieved an accuracy rate of 93% for Steenhoek et al. (2001) and 98% for Draganova et al. (2010a). Between the algorithms with the better results, it was chosen one to generate a model that was implemented and tested with program developed in this work. This model obtained a accuracy rate of 99,8% with this model. The best results were obtained with the combination of HSV color space, and texture characteristics with LBP pattern, and images with a resolution of 300 dpi.