Facial Expressions Images Generation Using Deep Convolutional Generative Adversarial Network (DCGAN)

Abstract

This research presents a DCGAN (deep convolutional generative adversarial network) for synthesizing grayscale facial expressions. The generation of grayscale facial expressions is essential for affective computing and emotion analysis. We presented a Deep Convolutional Generative Adversarial Network model designed for generating monochrome facial expression images, which were considered across two benchmark datasets, FER-2013 and CK+. The proposed implementation also benefits from non-linearity by performing Pixel Normalization and LeakyReLU activation to enhance image similarity and training reliability. As a result, the study achieved an FID of 88 and an IS of 3.0 on FER-2013, and an FID of 0.5 and an IS of 2.0 on CK+, which indicates high achieved image quality and photorealism. In a similar vein, the results estimated data quality and network robustness as well as their interaction with the GAN’s achieved performance. When taken together, the proposed research may serve as a strong lightweight facial expression synthesis baseline, which is likely to stimulate future GAN research and identity-boosting synthesis.