Normalized face image generation with perceptron generative adversarial networks

摘要

This paper presents a deep neural architecture for synthesizing the frontal and neutral facial expression image of a subject given a query face image with arbitrary expression. This is achieved by introducing a combination of feature space perceptual loss, pixel-level loss, adversarial loss, symmetry loss, and identity-preserving loss. We leverage both the frontal and neutral face distributions and pre-trained discriminative deep perceptron models to guide the identity-preserving inference of the normalized views from expressive profiles. Unlike previous generative methods that utilize their intermediate features for the recognition tasks, the resulting expression- and pose-disentangled face image has potential for several downstream applications, such as facial expression or face recognition, and attribute estimation. We show that our approach produces photorealistic and coherent results, which assist the deep metric learning-based facial expression recognition (FER) to achieve promising results on two well-known FER datasets.