Facial expression synthesis based on similar faces

Facial expression synthesis has several applications involving animation, human-computer interaction, entertainment, and training people with mental disorders. Facial expression synthesis aims to alter an image's facial expression, usually by reenacting the facial movements from an example image in the target image. Deformation-based approaches usually choose the example image manually, leading to different results depending on this choice. This study differs from the literature by proposing and evaluating techniques that consider the similarity between facial images to choose the source image. The primary goal is to investigate the influence of selecting the source image in generated facial expressions of emotions. We propose three techniques for selecting similar faces in the facial expression synthesis pipeline and compare them to other approaches. We also compare the generated synthetic emotions with the results of recent methods from the literature by using objective metrics. Our findings suggest that one of the proposed techniques presented higher results in the search for similar faces and similar or better results for the synthesis when compared to the literature. Additionally, a visual analysis showed that similar faces can improve synthetic images' realism, especially when compared to randomly selected facial images.