基于深度学习的单幅图片超分辨率重构研究进展  

A Review of Single Image Super-resolution Based on Deep Learning

作　　者：张宁[1,2] 王永成 张欣[1,2] 徐东东[1,2] ZHANG Ning;WANG Yong-Cheng;ZHANG Xin;XU Dong-Dong(Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033;University of Chinese Academy of Sciences,Beijing 100049)

机构地区：[1]中国科学院长春光学精密机械与物理研究所,长春130033 [2]中国科学院大学,北京100049

出　　处：《自动化学报》2020年第12期2479-2499,共21页Acta Automatica Sinica

基　　金：国家自然科学基金(11703027)资助。

摘　　要：图像超分辨率重构技术是一种以一幅或同一场景中的多幅低分辨率图像为输入,结合图像的先验知识重构出一幅高分辨率图像的技术.这一技术能够在不改变现有硬件设备的前提下,有效提高图像分辨率.深度学习近年来在图像领域发展迅猛,它的引入为单幅图片超分辨率重构带来了新的发展前景.本文主要对当前基于深度学习的单幅图片超分辨率重构方法的研究现状和发展趋势进行总结梳理:首先根据不同的网络基础对十几种基于深度学习的单幅图片超分辨率重构的网络模型进行分类介绍,分析这些模型在网络结构、输入信息、损失函数、放大因子以及评价指标等方面的差异;然后给出它们的实验结果,并对实验结果及存在的问题进行总结与分析;最后给出基于深度学习的单幅图片超分辨率重构方法的未来发展方向和存在的挑战.Super-resolution(SR)refers to an estimation of high resolution(HR)image from one or more low resolution(LR)observations of the same scene,usually employing digital image processing and machine learning techniques.This technique can effectively improve image resolution without upgrading hardware devices.In recent years,deep learning has developed rapidly in the image field,and it has brought promising prospects for single-image super-resolution(SISR).This paper summarizes the research status and development tendency of the current SISR methods based on deep learning.First,we introduce a series of networks characteristics for SISR,and analysis of these networks in the structure,input,loss function,scale factors and evaluation criterion are given.Then according to the experimental results,we discuss the existing problems and solutions.Finally,the future development and challenges of the SISR methods based on deep learning are presented.

关 键 词：深度学习 单幅图片超分辨率 卷积神经网络 生成对抗网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]