Super-resolution using texture guidance deep neural network

Abstract

Recently, convolutional neural network (CNN) based super-resolution (SR) algorithms have achieved substantal improvement on super-resolution task. However, the output of existing SR algorithms still have a problem with reconstructing fine texture regions. In this paper, we propose a texture guidance super-resolution deep neural network (TGSR) that uses learned texture feature maps for assis는ing super-resolution neural network to reconstruct elaborate details of corresponding high-resolution (HR) images from low-resolution (LR) images. A guidance of enhanced texture feature maps from texture reconstuction network make super-resolution network estimate more complex patterns or exquisite textures and improve the performance. Experimental results show that the proposed algorithm provides better performances in terms of PSNR and SSIM which is the measurement of SR task, compared to current state-of-the-art SR algorithms, while enhancing fine texture details.