Simultaneous Face Completion and Frontalization via Mask Guided Two-Stage GAN

Pose variation and occlusion are two key factors that affect the accuracy of face recognition. Most of the previous work alleviate the impacts of pose and occlusion by performing the tasks of face frontalization and face completion, respectively. Specially, generative adversarial networks (GANs) based methods have made a great progress on both of these two tasks. However, the two tasks are rarely paid attention simultaneously. Hence, the synthesis and recognition from the profile but occluded facial image is still an understudied and challenging problem in two aspects. 1) Occlusion mask, as a kind of noise, can be some very important prior information in the corrupted image. In particular, the occlusion mask is often used to fit this noise and help face restoration of the occluded region. However, a prior work, such as BoostGAN, failed to utilize the mask guided noise prior information. 2) The two tasks, de-occlusion and frontalization, are collaborative, so the identity discriminative information is easily lost if the two tasks are not organically unified in the training phase. In order to overcome these challenges, we propose a novel mask guided two-stage generative adversarial network (
TSGAN
). There are two major contributions in this work: 1) In order to utilize the prior information of noise, an occluded mask based attention model is proposed, which is integrated into the two stages via U-connection. This mask works as a guidance to simultaneously repair and frontalize the profile but occluded faces. 2) An end-to-end paradigm with a two-stage architecture (
i.e.,
face deocclusion network and face frontalization network) is proposed to complete the two tasks jointly. Furthermore, for preserving the discriminative identity information in both stages more effectively, we propose a novel dual triplet loss, consisting of a deocclusion triplet loss and a frontalization triplet loss. Qualitative and quantitative experiments on both constrained and unconstrained face datasets with regular and irregular (natural) occlusions demonstrate the superiority of our approach.