Emotion Recognition by the Facial Expressions using Machine Learning

Abstract

Being able to recognize emotions is a key component in communication between humans. This component however, is missing in human-machine communication. The recent advancements in Deep Learning (DL) brought with it many public data sets for Facial Expression Recognition (FER). Consequently, DL approaches became popular for emotion recognition, especially Convolutional Neural Networks (CNNs). In this work we present two different architectures utilizing single CNN for predicting basic human emotions. Furthermore, we present an architecture based on Decision Tree classifier, where the decision nodes are CNN models. Presented architectures are trained on a well-known FER data set (identified under the name of FER2013). We achieved state-of-the-art performance using a single CNN model based on the VGG architecture on this data set. We then evaluate performance of our model and test its generalization abilities over an external data set. In addition, we apply Transfer Learning methodology to reuse knowledge assimilated over FER2013 data set. We create a new model, that will be utilizing this transferred knowledge and test it over an external data set (identified under the name of CK+). Employing this new model, we achieved state-of-the-art performance on the CK+ data set.