Estimatição de eclosão em imagens de ovos do carrapato bovino baseado em redes neurais convolucionais

• Ano de defesa: 2023
• Autor(a) principal: SANTOS, Igor Silva
• Orientador(a): SILVA, Aristófanes Corrêa
• Banca de defesa: SILVA, Aristófanes Corrêa , COSTA JÚNIOR, Livio Martins , CAVALCANTE, André Borges , LOPES, Welber Daniel Zanetti
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO EM ENGENHARIA DE ELETRICIDADE/CCET
• Departamento: DEPARTAMENTO DE ENGENHARIA DA ELETRICIDADE/CCET
• País: Brasil
• Palavras-chave em Português: aprendizagem profunda; ovos férteis; processamento de imagem; carrapato; atenção suave;
• Palavras-chave em Inglês: deep learning; fertile eggs; image processing; tick; gentle attention.
• Área do conhecimento CNPq: Ciência da Computação; Engenharia Agrícola
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/4936

The cattle tick is an ectoparasite that causes losses of more than 3 billion dollars annually in cattle farming in Brazil, either by the transmission of diseases or by the reduction in the quality of the derived products. The use of chemical acaricides is the most common form of control. To choose the most effective acaricide, tests are carried out in the laboratory. Engorged females are used as samples and immersed in commercial solutions of different chemical classes. The parameters evaluated include weight of females, egg laying and the percentage of hatchability of larvae, which is determined by counting fertile and infertile eggs. This counting process is usually performed manually, which consumes a lot of time and is repetitive and tiring, and therefore, this dissertation aims to automate this procedure. In this context, a computational method is proposed to account for and estimate the percentage of hatchability based on image processing and deep learning techniques, which follows the flow: pre-pocessing, slide extraction, egg segmentation; classification and counting of eggs. The method proposes a convolutional neural network architecture with the inclusion of soft attention mechanisms, called DenseNetSA, which was compared with other network architectures. The method achieves promising results with the DenseNetSA network for the group with 6 images, with 98% of fertile eggs and 88.67% of infertile eggs correctly classified. For the group with 3 images, 98% of the fertile eggs and 90.3% of the infertile eggs were correctly classified. The percentage of hatching presented the following values: 96.35% ± 3.33; 95.98% ± 3.5; 0.0% ± 0.0 for the groups with three images in the Piracanjuba, Desterro and Barbalha populations, respectively; and 94.41% ± 3.84; 95.93% ± 2.36; 0.0% ± 0.0 for the groups with six images in the Piracanjuba, Desterro and Barbalha populations, respectively. There was no statistical difference between the evaluated methods. The automatic method for predicting the hatching percentage of R. microplus larvae was validated and proved to be effective, with considerable reduction in time to obtain results.