Convolutional Neural Networks Architectures for Facial Expression Recognition

This paper presents an overview of some convolutional neural networks architectures and methods for solving the facial expression recognition task using the FER_2013 image dataset. In this sense, three different convolutional networks have been trained and analyzed: a simple sequential architecture, a very lightweight network inspired from the XCEPTION architecture, and the well-known ResNet50 model. Following the experiments conducted using the above-mentioned architectures we have managed to obtain a state-of-the art comparable, 71.25% prediction accuracy on the test subset of FER-2013.