Facial Attribute Editing using Semantic Segmentation
Peng Chen, Qi Xiao, Jian Xu, Xiaoli Dong, Linjun Sun · 2019
Recently, great success has been made in facial attribute editing by using adversarial learning. However, most existing studies focus on how to generate an image with high fidelity while ignore where to manipulate. Hence, in this paper, generative adversarial network with semantic masks (SM-GAN), a new framework for accurate facial attribute editing is proposed. Specifically, the proposed framework is constructed by combining GAN with semantic segmentation network. Here, the semantic segmentation network provides the mask with respect to attribute-related region so that the editing only occurs in expected areas. Qualitative and quantitative experiments on public dataset CelebA demonstrate that the proposed approach can not only generate realistic attribute editing results, but also preserve attribute-irrelevant areas unchanged.