Detecção de árvores em imagens aéreas com métodos de saliência e redes neurais profundas

• Autor(a) principal: Sousa, Naiane Maria de
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFG
• Texto Completo: http://repositorio.bc.ufg.br/tede/handle/tede/11838
• topic: Agricultura; Visão computacional; Classificação de objetos; Aprendizado profundo; Agriculture; Computer vision; Object classification; Deep learning; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

Precision agriculture grows along with new technologies. The use of images for agricultural analysis has been growing and has low cost and processing agility. The presence of trees, with green foliage, inside or on the edges of crops can influence the results of the observation of plantations. The color of the tree tops can be confused with the color of the cultivar's leaves. In this work, we intend to assist in the detection of trees in aerial agricultural images using computational saliency as support. A literature review was conducted, that confirmed the applicability of the saliency methods in different sectors of agriculture. Then, experiments were carried out to evaluate the salient condition of trees in the studied context. It was noticed that some saliency methods highlight trees near to crops, however other regions are also highlighted. Neural networks were used to classify objects obtained from salient regions as being a tree or not. The classifiers achieved about 96\% accuracy. The Tree Detection Method created during this research has the potential to identify regions of trees present in a crop area, through aerial images. The proposed method, using the ResNet classifier, was be able to find about 69\% of the trees in the sample test images.