| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 17.96 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Recentemente, detectou-se em Portugal, o inseto Trioza erytrea, o qual tem a possibilidade
de transmitir a doença Huanglongbing (Citrus Greening) prejudicial à cultura dos citrinos.
Esta patologia ainda não possui cura e tem alto potencial lesivo ao cultivo. Embora
a doença não se manifestou-se em território português, métodos preventivos, baseados no
monitoramento da quantidade de Triozas, são indispensáveis para efetivas estratégias de
contingência, assim como antecipar uma possível disseminação. O método tradicional de
monitoramento é realizado por intermédio de um profissional incumbido de identificar e
contabilizar manualmente a Trioza e mais quatro outros insetos (indicadores da presença
de Trioza), a partir de armadilhas autocolantes instaladas nas plantações de citros. Esse
modo de execução consome tempo considerável, é suscetível a erros e demanda um profissional
que nem sempre está disponível. Nesse sentido, o presente trabalho se propôs a
desenvolver uma solução computacional, mediante a utilização de Redes Neurais Convolucional
(RNC). Isso se deu por meio de uma aplicação, a qual desempenha a função de
identificar, classificar e contar automaticamente os insetos citados. Realizou-se o estudo
do desempenho de arquiteturas de RNC, as quais possuem a potencialidade de reconhecer
insetos a partir de imagens. Como resultado, mostrou-se que a arquitetura SSD com
Inception-v2 obteve melhor desempenho na identificação de Trioza. Porém, num primeiro
momento, não se tem a viabilização da substituição do método tradicional pela aplicação,
isso em decorrência das deficiências nos processos de aprendizagem.
Recently, the insect Trioza erytrea was detected in Portugal, which has the possibility of transmitting the disease Huanglongbing (Citrus Greening) harmful to the citrus culture. This pathology has no cure yet and has a high potential for injury to crops. Although the disease did not manifest itself in Portuguese territory, preventive methods, based on monitoring the amount of Triozas, are indispensable for effective contingency strategies, as well as anticipating a possible spread. The traditional method of monitoring is carried out by means of a professional responsible for manually identifying and accounting for Trioza and four other insects (indicators of the presence of Trioza), using self-adhesive traps installed in citrus plantations. This mode of execution consumes considerable time, is susceptible to errors and demands a professional who is not always available. In this sense, the present work proposed to develop a computational solution, through the use of Convolutional Neural Networks (CNN). This was done through an application, which performs the function of automatically identifying, classifying and counting the insects mentioned. The study of the performance of CNN architectures was carried out, which have the potential to recognize insects from images. As a result, it was shown that the SSD architecture with Inception-v2 achieved better performance in identifying Trioza. However, at first, there is no possibility of substituting the traditional method for application, due to deficiencies in the learning processes.
Recently, the insect Trioza erytrea was detected in Portugal, which has the possibility of transmitting the disease Huanglongbing (Citrus Greening) harmful to the citrus culture. This pathology has no cure yet and has a high potential for injury to crops. Although the disease did not manifest itself in Portuguese territory, preventive methods, based on monitoring the amount of Triozas, are indispensable for effective contingency strategies, as well as anticipating a possible spread. The traditional method of monitoring is carried out by means of a professional responsible for manually identifying and accounting for Trioza and four other insects (indicators of the presence of Trioza), using self-adhesive traps installed in citrus plantations. This mode of execution consumes considerable time, is susceptible to errors and demands a professional who is not always available. In this sense, the present work proposed to develop a computational solution, through the use of Convolutional Neural Networks (CNN). This was done through an application, which performs the function of automatically identifying, classifying and counting the insects mentioned. The study of the performance of CNN architectures was carried out, which have the potential to recognize insects from images. As a result, it was shown that the SSD architecture with Inception-v2 achieved better performance in identifying Trioza. However, at first, there is no possibility of substituting the traditional method for application, due to deficiencies in the learning processes.
Description
Keywords
Aprendizagem profunda Redes neurais convolucionais Citros Processamento de imagens Trioza erytre Faster R-CNN SSD Inception v2 ResNet-50