Sistema de reconhecimento de expressões faciais
Facial Expressions give out a lot of information on emotions and social interactions of individuals. Systems capable of detecting facial expressions could be used for an wide range of applications, such as: fatigue detection, suspicious activity detection, cellphone integration, learning environment...
| Principais autores: | Zanellato, André Lucas, Ferreira, Cristian Bortolini |
|---|---|
| Formato: | Trabalho de Conclusão de Curso (Graduação) |
| Idioma: | Português |
| Publicado em: |
Universidade Tecnológica Federal do Paraná
2020
|
| Assuntos: | |
| Acesso em linha: |
http://repositorio.utfpr.edu.br/jspui/handle/1/8161 |
| Tags: |
Adicionar Tag
Sem tags, seja o primeiro a adicionar uma tag!
|
| Resumo: |
Facial Expressions give out a lot of information on emotions and social interactions of individuals. Systems capable of detecting facial expressions could be used for an wide range of applications, such as: fatigue detection, suspicious activity detection, cellphone integration, learning environments, entertainment and client intearction. This project consists of developmenting a system capable to recogninize and classify facial expression presented in an frontal image of an human face, resulting in one of the 7 facial expressions: happiness, sadness, fear, anger, surprise, disgust and neutral. In order to do this task, the method choosen was the training of a Deep Neural Network made of two layers: a Convolutional Neural Network, capable of extraction features from an image, and a Fully-Connected Neural Network in charge of the classification. Both networks were implemented with the help of the Caffe framework, developed by Berkeley University, avaliable in Python and C++ and trained with the extended Cohn-Kanade database. For said database, it was necessary to perform database augmentation, since Deep Neural Networks require a lot of data. Before training with the 7 facial expressions, there was a preliminary training with only three, to test the architetcure. The choosen architecture was the winner of the 2012 ImageNet Challange, called Alexnet, modified slightly to better fit our data. After results that supported the hipothesis of a good result, we did the training with the 7 expressions. The obtained results are displayed in confusion matrixes, which allows the analysis of the classifications in cases where they were right or wrong. In the end, we implemented a system which uses the neural network that we trained to recognize expressions in real time. |
|---|