ABSTRACT
Detecting facial expressions is a vital aspect of interpersonal communication. Automatic facial emotion recognition (FER) systems for detecting and analyzing human behavior have been a subject of study for the past decade, and have played key roles in healthcare, crime detection, and other use cases. With the worldwide spread of the COVID-19 pandemic, wearing face masks while interacting in public spaces has become recommended behavior to protect against infection. Therefore, improving existing FER systems to tackle mask occlusion is an extremely important task. In this paper, we analyze how well existing CNN models for FER fare with masked occlusion and present deep CNN architectures to solve this task. We also test some methods to reduce model overfitting, such as data augmentation and dataset balancing. The main metric used to compare the models is accuracy, and the dataset used here is FER2013. Images from FER2013 were covered by masks using a certain transformation, resulting in a new dataset, MFER2013. From our evaluation and experimentation, we found that existing models need to be modified before they can achieve good accuracy on masked datasets. By improving the architecture of the base CNN, we were able to achieve a significantly improved accuracy. © 2022 IEEE.