FSRGAN-DB: Super-resolution Reconstruction Based on Facial Prior Knowledge

Abstract

Face super-resolution (SR) reconstruction is a method of reconstructing a high-resolution (HR) face image from a low-resolution (LR) face image with more facial details. How-ever, most of the SR methods do not account for facial structures and suffer from loss of face details. In this paper, we propose a method that explicitly incorporates structural information of faces into the face super resolution network. We firstly use Super-Resolution Generative Adversarial Network(SRGAN) as the basic network and improve it by replacing residual blocks with dense blocks(SRGAN-DB). Then, we use the heat map of facial key points to describe the facial structure and the trend of facial contours. When the resolution of low-resolution images is very low, it is difficult to obtain fine facial features. But it is easy to get the information of facial key points. Specifically, the LR image is sent to two network branches, one obtains the fine features of the image, and the other generates a heat map of the facial key point positions. In order to enhance the feature expression of facial key points information, we convert the facial key points heat map into a binary image and connect it to the face fine encoder network. Extensive experiments prove that the proposed method is superior to existing network methods to face super-resolution reconstruction.