Face Super-Resolution with Better Semantics and More Efficient Guidance

Abstract

Recently, facial priors have been widely used to improve the quality of super-resolution (SR) facial images, but it is underutilized in existing methods. On the one hand, facial priors such as semantic maps may be inaccurately estimated on low-resolution (LR) images or low-scale feature maps with $L_{1}$ loss. On the other hand, it is inefficient to guide SR features with constant prior knowledge via concatenation at only one intermediate layer of the guidance network. In this paper, we focus on face super-resolution (FSR) based on semantic maps guidance and propose two simple and efficient designs to address the above two limitations respectively. In particular, to address the first limitation, we propose a novel one-hot supervision strategy to pursue accurate semantic maps, which focuses more on penalizing misclassified pixels by relaxing the regression constraint. In addition, a semantic progressive guidance network (SPGN) is proposed that uses semantic maps to learn modulation parameters in normalization layers to efficiently guide SR features layer by layer. Extensive experiments on two benchmark datasets show that the proposed method improves the state-of-the-art in both quantitative and qualitative results at $\times$8 scale.