深度学习单帧图像超分辨率重建研究综述  

Survey of single image super-resolution reconstruction with deep learning

作　　者：曲延云[1] 陈蓉 李翠华[1] 王菡子[1] QU Yanyun;CHEN Rong;LI Cuihua;WANG Hanzi(School of Informatics,Xiamen University,Xiamen 361005,China;Information Engineering College,Xizang Minzu University,Xianyang 712082,China)

机构地区：[1]厦门大学信息学院,福建厦门361005 [2]西藏民族大学信息工程学院,陕西咸阳712082

出　　处：《厦门大学学报（自然科学版）》2021年第3期555-570,共16页Journal of Xiamen University：Natural Science

基　　金：国家自然科学基金(61876161,U1065252);西藏自治区自然科学基金(XZ2018ZRG-65)。

摘　　要：单帧图像超分辨率重建(single image super-resolution reconstruction,SISR)是一类典型的不适定问题,具有广泛的实际应用价值.近年来,基于深度学习的SISR方法的研究已经取得长足的进展,其性能大大超越了传统的方法.本文将SISR研究归纳为两大挑战性问题:SISR的建模和面向真实场景的领域自适应问题.基于此,将当前基于深度学习的SISR方法划分为有监督的SISR和不成对的SISR.进一步简单介绍SISR的应用情况.最后,总结并探讨未来SISR研究的发展方向.As a typical ill-posed problem,single image super-resolution reconstruction(SISR)has a wide range of practical applications.In recent years,SISR methods based on the deep learning have made great progress,and outperform traditional methods.This survey first summarizes challenging problems in SISR,such as modeling SISR via deep learning and adapting to the real world.Next,we divide the existing SISR methods based on the deep learning into supervised and unpaired.Then,the application of SISR is briefly introduced.Finally,conclusions are drawn and future research trends are discussed.

关 键 词：单帧图像超分辨率重建 深度学习 有监督学习 不成对学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]