基于生成对抗网络的无监督图像超分辨率算法  

Unsupervised image super-resolution algorithm based on Generative Adversarial Network

作　　者：赵志博 滕奇志[1] 任超[1] 何小海[1] 翟森 Zhao Zhibo;Teng Qizhi;Ren Chao;He Xiaohai;Zhai Sen(College of Electronics and Information Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电子信息学院,四川成都610065

出　　处：《信息技术与网络安全》2022年第1期55-62,共8页Information Technology and Network Security

基　　金：国家自然科学基金(62171304,61801316)。

摘　　要：目前,大多数基于学习的图像超分辨率研究通常采用预定的降质类型(比如双三次下采样)处理高分辨率图像,来产生成对的训练集。然而,真实图像往往存在未知的模糊和噪声,导致这些算法无法有效应用到真实场景中。为了实现真实图像的超分辨率重建,提出了一种基于生成对抗网络的无监督图像超分辨率算法,所提出的算法分为域转换子网络和重建子网络两个部分。同时设计了深度特征提取模块,通过融合不同感受野所提取的图像特征来提升网络的性能。实验结果证明,相比于目前多数的图像超分辨率算法,本文算法能够实现真实降质图像(存在噪声、模糊等)的图像超分辨率,在主观效果和客观指标上均能获得更好的性能。In most existing researches on learning-based image super-resolution,the pair of training datasets is generated by down-scaling high-resolution(HR)images through a predetermined operation(e.g.,bicubic down-sampling).However,these algorithms cannot be effectively applied to real scenes since the real-world image contains unknown noise and blur.To this end,we propose an unsupervised image super-resolution algorithm based on Generative Adversarial Network in this paper.Our method contains two parts:domain conversion sub-network and reconstruction sub-network.In addition,the deep feature extraction module is proposed to improve the performance of the network by merging the image features captured by different receptive fields.Extensive experiments illustrate that compared with most current image super-resolution algorithms,the proposed method can be applied to real-world image(containing noise,blur,etc.)super-resolution,and achieves the start-of-the-art(SOTA)performance on both subjective and objective evaluations.

关 键 词：真实图像超分辨率 域转换 生成对抗网络 无监督训练 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391[自动化与计算机技术—控制科学与工程]