Models for Real-Time Emotion Classification: FER-2013 Dataset

The recognition of emotions and facial expressions is a topic of study that is intertwined with several fields of research: for example, artificial intelligence, digital games, and human-computer interaction (HCI). Head-mounted displays (HMD) have revolutionized virtual worlds. These devices have allowed people to interact in environments with great realism. Recognizing emotions and facial expressions within an HMD has been a great challenge because they cause optical occlusions. The same HMDs occlude important areas of the face, such as the eyes. However, as of 2021, some of the latest HMDs manufactured have incorporated cameras and/or sensors to control what is happening behind these devices. Our working group is developing a project whose objective is to develop and validate a set of techniques that allow us to recognize emotions and facial expressions so that they can be projected onto an avatar in real-time within these HMDs. This work, which is part of the first stage of the project, documents the results of three CNN models designed to recognize emotions in real-time. We adapted three CNN structures for training on the well-known FER-2013 dataset. Our three models are made up of one, two, and three convolutional channels respectively and work by dividing the input image for analysis without any other preprocessing. According to our objectives, the proposed models achieved promising results in terms of accuracy values and training time. One of the models reaches 63.0% accuracy, with a standard deviation of ± 0.005, considering that no preprocessing is performed on the input images.