Using Generative Adversarial Networks for the synthesis of emotional facial expressions in virtual educational environments

Abstract

The generation of emotional facial expressions using Generative Adversarial Networks (GANs) has been widely researched, achieving significant advances in creating high-quality images. However, one of the main challenges remains the accurate transmission of complex and negative emotions, such as anger or sadness, due to the difficulty of correctly capturing the facial micro gestures that characterize these emotions. Traditional GAN architectures, such as StyleGAN and DCGAN, have proven highly effective in synthesizing positive emotions such as joy. However, they have limitations regarding more subtle emotions, leading to a higher rate of false negatives. In response to this problem, we propose fine-tuning the GAN discriminator. This tuning optimizes the discriminator’s ability to identify the minor details in facial expressions by using perceptual losses, allowing for better differentiation between the generated emotions, reducing classification errors, and improving precision in difficult-to-represent emotions. The results obtained in this study demonstrate a significant improvement in the system’s precision, especially in complex emotions. Precision for the anger emotion increased from 85.7% to 89.1%, and the number of false negatives was reduced from 16 to 10. Overall, precision for complex emotions exceeded 85%, substantially improving traditional solutions. These results demonstrate the potential of fine-tuning the GAN architecture for applications requiring more faithful and effective emotional interaction, significantly improving user experience across multiple domains.