Progressive face super-resolution via learning prior information

Abstract

Face super-resolution is a branch of the field of super-resolution. It is mainly aimed at the reconstruction of face images and distinguishing SR from general images. Face geometry prior information is used to optimize the face super-resolution network, which can generate high-resolution face images with better visual quality from low-resolution face images. In order to further improve the visual quality of reconstructed images, an improved face super-resolution network is proposed. In this paper, the key module of FSRNet is improved and new loss function is added to achieve a better face super-resolution network. Specifically, our job is to:(1) In the generated rough SR face image, we input low-resolution(16 x16) face image, then use the Deconv convolution enlarge images.(2)By introducing heatmap loss, facial attention loss and adversarial training to reduce the artifacts of FSRNet network. (3) We divide the network into two-step training, first train coarse SR network, get SR images quickly, and then we train the rest of the network. The final output is a super-resolution face with high visual quality.