Desenvolvimento e validação de um sistema de identificação de emoções por visão computacional e redes neurais convolucionais com transferência de aprendizado

• Ano de defesa: 2017
• Autor(a) principal: Matos, Hamilton de
• Orientador(a): Notargiacomo, Pollyana Coelho da Silva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Presbiteriana Mackenzie
• Palavras-chave em Português: expressões faciais; visão computacional; aprendizagem de maquinha; redes neurais convolucionais; transferência de aprendizado
• Área do conhecimento CNPq: CNPQ::ENGENHARIAS
• Link de acesso: http://dspace.mackenzie.br/handle/10899/24469

This research aims to establish a methodology for the creation of a convolutional neural network in situations where there is a reduced number of data for training. The scenario adopted is the identi_cation of facial expressions in images and their respective emotions. Through the validation of the proposed method, in addition to the availability of the methodology itself, a public access interface is also available for providing information about the emotions detected in an image. In this context, human behavior studies such as Paul Ekman's analysis and definitions of universal emotions are used. Diferent techniques of computer vision and machine learning with convolutional neural networks were used to create the software. The aim of this work is to verify the possibility of real-time analysis, aiming to provide data in situations in which feedback can be fundamental for making certain decisions (eg in reactions in statements, for example). Even on a reduced training data set, it was possible to achieve an average final test accuracy of 65 % with an average training time of ten minutes.