Guided Super-Resolution Restoration of Single Image Based on Image Quality Evaluation Network

Abstract

SISR (Single image super-resolution) has always been a key problem in image processing field. In recent years, deep learning has been successfully used to SISR reconstruction. However, most of the previous deep learning methods use L2 norm based on pixel pairs as loss function, which results in a high peak signal-to-noise ratio (PSNR) value, but the perception quality has not been improved. When using Generative Adversarial Network (GAN), although it has good perception quality, PSNR is lower. So we’ll generate realistic results when both of them are used well. The image quality evaluation (IQA) network is to evaluate the image quality, so as to obtain good PSNR value and perception quality. In this paper, we use image quality assessment network to guide the SISR reconstruction network. Besides that, our proposed Super-resolution reconstruction of single image method is composed of several our given cross-attention units (CA) and is trained iteratively. Experimental results demonstrate that our method in qualitative and quantitative is better than others.