Classificação automática do Diaphorina citri em imagens de microscopia

• Ano de defesa: 2016
• Autor(a) principal: Melo, José Leonardo dos Santos
• Orientador(a): Angelo, Michele Fúlvia
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual de Feira de Santana
• Programa de Pós-Graduação: Mestrado em Computação Aplicada
• Departamento: DEPARTAMENTO DE TECNOLOGIA
• País: Brasil
• Palavras-chave em Português: Diaphorina Citri; Armadilhas adesivas amarelas; Aprendizagem de máquina
• Área do conhecimento CNPq: CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO
• Link de acesso: http://localhost:8080/tede/handle/tede/377

The Huanglongbing (HLB) is the disease of greatest concern for growers because they spread quickly and cause severe symptoms. The Diaphorina citri insect is the main vector of the HLB. The application of insecticides is a control measure of the vector insect of the HLB widely adopted. The amount of pesticides needed for an effective control of this insect is better estimated if such application is combined with a monitoring of its population by yellow sticky traps. These insects are captured for a manual count in research centers. So, this research aims to discover a computational approach of classification of Diaphorina citri insect images with higher accuracy rate that the classification rate currently used in manual counting procedure and thus enable the automation of this important counting procedure. For this, have been tried and combined computational methods for features extraction (ORB, SIFT, SURF, BRISK and FREAK), grouping of characteristics (Mini Batch K-Means) and features classification for machine learning (KNN and SVM), using a generated bank with 1152 images of insects. The best found classification approach (extractor SURF/SIFT, BoF with Diaphorina citri features and SVM with core RBF) generated classification performance results for the metric accuracy, which outperformed the best measured result in research that evaluated the counting manual process. In this approach, the highest achieved accuracy, in the cross validation process, was 98.17% and was 2.54% as standard deviation and the accuracy of the final test of generalization model was 99.14%. The achieved result is of great importance for the control of HLB. The achieved classification accuracy rates were higher than rates reported in the manual procedure, making possible the construction of computer systems to high accuracy for the control of this insect. This automated control can provide significant savings of funds.