Deep learning na identificação e quantificação de flores de Catharanthus roseus (L.) G. Don

• Ano de defesa: 2022
• Autor(a) principal: Carmélia Maia Silva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ICA - INSTITUTO DE CIÊNCIAS AGRÁRIAS; Programa de Pós-Graduação em Produção Vegetal; UFMG
• Palavras-chave em Português: Vinca; Quantificação de flores; Visão Computacional; Mask R-CNN; Análise de imagem; Catharanthus; Arquitetura paisagística; Jardins
• Link de acesso: http://hdl.handle.net/1843/46536

Catharanthus roseus (L.) G. Don popularly known as vinca, is a medicinal and ornamental species, in landscaping it is an option for the composition of the garden, it has a flowering of different colors, in addition to be resistant to regions with a tropical climate. Computer vision in the landscape sector presented tools to help in the development and management of landscape projects such as applications for the identification of plants, but for the identification and quantification of flowers through images, there is still little information. Thus, the objective of this work was to evaluate the potential of the Mask R-CNN network to quantify vinca flowers and qualify them in terms of color for application in landscape projects. 700 images were collected, 500 had both pink and white flowers and 200 had just the leaves that made up the background. For the composition of the computer-generated image bank, 100 white flowers and 100 roses were processed in png format and formed the foreground, the two being separated into two subclasses. The training using the transfer learning technique with the Mask R-CNN algorithm was carried out in collaborative Google, with commands in python language and libraries from the github platform. Using quality raters, the Mask R-CNN convolutional neural network showed an accuracy of 97% for the pink vinca subclass and 83% for the white vinca subclass, and an overall accuracy greater than 80%. The network proved effective in estimating the number of flowers, in addition to detecting and segmenting them, qualifying them in terms of color. Therefore, the methodology can be used in landscaping to estimate and quantify flowers through images for garden composition.