Super-resolving Tiny Faces with Face Feature Vectors

Abstract

Most of the current state-of-the-art tiny face super-resolution (SR) methods aim at learning a single one-to-one mapping to super-resolve low-resolution (LR) face images. In contrast with high-resolution (HR) faces images, LR faces images lack fine facial details, implying that an LR face image can be mapped to many HR candidates or vice versa. This ambiguity may lead that an HR face image super-resolved by one-to-one SR methods cannot preserve the accurate facial details. To alleviate this problem, we consider tiny face SR as an one-to-many mapping, and demonstrate that injecting reasonable additional facial prior knowledge can significantly reduce the ambiguity in face SR. Specifically, with the GAN architecture, we propose a novel face SR network consisting of an upsampling network and a discriminative network. The upsampling network is designed to embed facial prior knowledge (represented as a vector) into the residual features of LR inputs and super-resolve LR inputs $(16 \times 16$ pixels) with up-scaling factor of $4 \times$. The discriminative network aims at examining whether the generated HR face images matches the corresponding facial prior. Furthermore, we also propose a model which can map the samples of a standard normal distribution to the facial prior distribution. In this way, we can readily sample different facial prior to support us to super-resolve a single LR face image to abundant different HR faces images. Through conducting extensive evaluations on a large-scale dataset, we demonstrate that our method achieves encouraging results, outperforming the current competitive algorithms.