Generative Adversarial Network for Facial Emotion Recognition: A Feasibility Study

Abstract

Integration of artificial intelligence into different domains has been a trending topic over the past few years. A closed-loop feedback system which immerses the subject in a virtual reality environment with a novel reward platform is being developed to help people suffering from autism spectrum disorder. In this work, the feasibility of using generative adversarial networks to generate synthetic images by restructuring unseen input data to match that of the training set for the recognition of human emotions is being studied. System performance was based on true positive predictions from the different classification models developed in previous work. Preliminary results showed that the proposed system was able to improve class predictions, but lacked in the ability to generate different class sets. The performance highlights the feasibility of this method and its practical applications in generating more data and improving model robustness.