A Two-Tier GAN Architecture for Conditioned Expressions Synthesis on Categorical Emotions

摘要

Abstract Emotions are an effective communication mode during human–human and human–robot interactions. However, while humans can easily understand other people’s emotions, and they are able to show emotions with natural facial expressions, robot-simulated emotions still represent an open challenge also due to a lack of naturalness and variety of possible expressions. In this direction, we present a two-tier Generative Adversarial Networks (GAN) architecture that generates facial expressions starting from categorical emotions (e.g. joy, sadness, etc.) to obtain a variety of synthesised expressions for each emotion. The proposed approach combines the key features of Conditional Generative Adversarial Networks (CGAN) and GANimation, overcoming their limits by allowing fine modelling of facial expressions, and generating a wide range of expressions for each class (i.e., discrete emotion). The architecture is composed of two modules for generating a synthetic Action Units (AU, i.e., a coding mechanism representing facial muscles and their activation) vector conditioned on a given emotion, and for applying an AU vector to a given image. The overall model is capable of modifying an image of a human face by modelling the facial expression to show a specific discrete emotion. Qualitative and quantitative measurements have been performed to evaluate the ability of the network to generate a variety of expressions that are consistent with the conditioned emotion. Moreover, we also collected people’s responses about the quality and the legibility of the produced expressions by showing them applied to images and a social robot.