FAE-GAN: facial attribute editing with multi-scale attention normalization

摘要

Facial attribute editing has gained increasing attention recently. Previous methods tackle this challenge by incorporating encoder-decoder and generative adversarial networks. However, the bottleneck layer in encoder-decoder of these methods often leads to blurry and low-quality editing results. And skip connections are used between deep and shallow layers to improve image quality but suffer from a limited ability to manipulate attribute. To address these issues, we propose a novel Facial Attribute Editing Generative Adversarial Networks from a selective refinement perspective, which is capable of focusing on editing the image attributes to be changed while preserving its unique details. Specifically, our method first learns a spatially varying function that maps a high-level feature map to an appropriate parameter map of the normalization layer. Then, by utilizing the residual block, the low-level feature map is added to the feature map after modulation, making the attribute refinement task easier. Experimental results show the superiority of our method in both performances of the attribute manipulation accuracy and perception quality.