Facial Expression Translation using Cycle Consistent Adversarial Networks with Contrastive Loss

Altering facial expressions in images using artificial intelligence can be a complex and challenging task. It is crucial to maintain the original identity and features of the face while modifying its expression. AI-based image generation technology requires a dataset to generate realistic images. However, this process can sometimes result in losing or modifying the original face's identity during expression alteration. To address this issue, we propose using cycle-consistent adversarial networks (CycleGAN) with contrastive loss to translate facial expressions. Our experiment utilizes the RaFD dataset, and the results presented in this paper only showcase the translation of neutral-to-happy and contempt-to-happy expressions. The generated output achieved an FID of 72.731 and a KID of 0.049, with an FSD of 0.548 for neutral-to-happy expression translation, and an FID of 83.665 and a KID of 0.052 with an FSD of 0.495 for happy-to-neutral expression translation.