Masked FER-2013: Augmented Dataset for Facial Expression Recognition

Abstract

This poster introduces the Masked PER -2013 dataset that can be used to analyze facial emotions in people wearing masks and train a recognizer. The PER-2013 dataset containing face images annotated with seven emotions was modified by synthesizing mask images into the lower portion of the face. Based on the quantitative evaluation, our dataset can improve the accuracy of CNN, VGG, and ResNet models' emotion recognition in masked images by a maximum of 46%. Additionally, as a use case, we present an application that recognizes the emotions of a masked person, generates emotional face animation, and augments it on a real mask during video conferencing.