Deep learning na análise dos testes de tetrazólio e de pureza física de sementes de Urochloa brizantha cv. Marandu

• Ano de defesa: 2021
• Autor(a) principal: Rayane Aguiar Alves
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ICA - INSTITUTO DE CIÊNCIAS AGRÁRIAS; Programa de Pós-Graduação em Produção Vegetal; UFMG
• Palavras-chave em Português: Sementes; Qualidade; Capim-marandu; Deep learning; Redes neurais convolucionais
• Link de acesso: http://hdl.handle.net/1843/36081

The computational image analysis associated with the Mask R-CNN (Mask Region-based Convolutional Neural Network) can be used to automate and improve several routine evaluations, such as the tetrazolium test and seed purity analysis. The tetrazolium test is considered one of the fastest tests in the evaluation of seed viability, being highly dependent on the experience and ability of the evaluator to classify the results, thus being a subjective test. The analysis of physical purity is a mandatory determination for commercialization of seeds and indicates the physical quality of lots. Thus, the objective was to verify the efficiency of computational automation in the analysis of the tetrazolium and physical purity tests of Urochloa brizantha cv. Marandu through the Mask R-CNN. The experiment was conducted in Montes Claros. Batches of Brachiaria seeds from commercial locations were used. To obtain the images, in both tests, a flatbed scanner (Hp Officejet 4500 Desktop) was used. The tetrazolium test was performed according to the Seed Analysis Rules (0.1% solution, for 5 hours, at 40 ºC) (BRASIL, 2009). Each of the two fragments of the seeds was placed on the scanner glass plate with the colored part facing downwards. To automate the classification of the results of the tetrazolium test, 4 repetitions per lot were used and each repetition had 50 seeds. As for the automation of the purity analysis, 5 purity proportions were scanned, each proportion having two samples of 10 grams each. The neural networks showed high efficiency when classifying the results of the tetrazolium test, presenting accuracy values above 80%, and general accuracy around 91%. It also obtained efficiency in the estimation of the physical purity of the evaluated lots, with high coefficients of determination. Therefore, the automation of the analysis of tetrazolium tests and purity analysis in seeds of Urochloa brizantha cv. Marandu is viable using the Mask R-CNN technique.