Detecção de anemia em ovinos através de aprendizagem profunda em imagens de mucosa ocular

• Ano de defesa: 2021
• Autor(a) principal: Almeida, Antônio Márcio Albuquerque
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: FAMACHA; Detecção de anemia e ovinos; Segmentação; Classificação; Aprendizagem Profunda
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/73133

The FAMACHA method is used as an instrument for parasitological diagnosis, and it can be applied to sheep to detect different degrees of anemia. This is intended to assist in the application of dewormer only for sick animals, but it is still a method that depends on human interpretation. The technology has helped in the process of automatic and accurate disease detection. In this work, we performed the segmentation and classification of images of the mucosa of the eye of sheep for the detection of anemic animals based on deep learning models. This work used a total of 106 distinct images of the ocular mucosa of sheep, collected in the environment of EMBRAPA - Goats and Sheep, located in the region of Sobral - Ceará. This dataset was separated into training set, validation set, and test set. Using several models for segmentation, it was possible to find the region of interest in all these images, and the best model obtained an accuracy of 97.29 % in the test set, for the Jaccard similarity method. For the classification procedure, the best model obtained an accuracy of 95.23 % in detecting the animal’s anemia status. This result was obtained from the validation set, with a total of 21 images. With the segmentation and classification models already trained, tests were performed in a mobile application, in which the criterion was to seek the shortest processing time, for segmentation it was 0.375 (s) and for classification it was 0.121 (s). With the samples also collected in another database made available by researchers from UFMA, in which its best result for segmentation was 78.76 %, in the Jaccard similarity method, while for classification its accuracy was of 64.76%.