|
Acadêmico(a): DIÓGENES ADEMIR DOMINGOS |
Título: DEEP-EMOTIVE: PROTÓTIPO DE SISTEMA PARA RECONHECIMENTO DE EXPRESSÕES FACIAIS UTILIZANDO APRENDIZADO PROFUNDO |
|
Abstract: |
This work presents the development of a prototype for emotion recognition of the Ekman et al. (1987) universal facial expressions using deep learning. It was applied convolutional neural networks to extract the facial features, which is the state of the art technique in computer vision area, used on the object recognition of images. To the development of this work it was used Digital Image Processing techniques jointly with the Keras framework, to build the deep neural networks. The Cohn-Kanade AU-Coded Expression Database Version 2 (CK +) was applied to train and test the prototype. The proposed work achieved all the pre- established objectives, being able to recognize eight emotions: joy, sadness, fear, surprise, disgust, anger, contempt and neutral through the extracted features of facial expressions. The prototype reached a 96.33 % of accuracy, demonstrating the feasibility of the techniques used. In addition, the transfer learning technique validated the algorithm developed in which it was tested in two new databases: Japanese Female Facial Expression (JAFFE), obtaining a precision of 93.02%, and Facial Expression Recognition 2013 (FER-2013), with an accuracy of 60.62%. The development of this prototype proved to be relevant on the support of new computational algorithms based on this architecture to the emotion recognition of facial expressions. |
|
|