近年来超分结果对比

近几年图像超分辨率经典方法结果对比

BI degradation：bicubic退化，即将HR图像进行双三次插值，缩下S倍，生成LR图像。

Quantitative results with BI degradation model. Best and second best results are highlighted and underlined

Method	Scale	params	Set5	Set14	BSD100	Urban	Manga109
EDSR	x2	40.7M	38.11/.9602	33.92/.9195	32.32/.9013	32.93/.9351	39.10/.9773
RCAN	x2	15.4M	38.27/.9614	34.14/.9216	32.41/.9027	33.34/.9384	39.44/.9786
SAN	x2	-	38.31/.9620	34.07/.9213	34.42/.9028	33.10/.9370	39.32/.9792
CSNLN	x2	-	38.28/.9616	34.12/.9223	32.40/.9024	33.25/.9386	39.37/.9785
RFANet	x2	-	38.26/.9615	34.16/.9220	32.41/.9026	33.33/.9389	39.44/.9783
HAN	x2	-	38.27/.9614	34.16/.9217	32.41/.9027	33.35/.9385	39.46/.9785
EBPN	x2	-	38.35/.9620	34.24/.9226	32.47/.9033	33.52/.9402	39.62/.9802
IRN	x2	1.66M	43.99/.9871	40.79/.9778	41.32/.9876	39.92/.9865	-
EDSR	x3	-	34.65/.9280	30.52/.8462	29.25/.8093	28.80/.8653	34.17/.9476
RCAN	x3	-	34.74/.9299	30.65/.8482	29.32/.8111	29.09/.8702	34.44/.9499
SAN	x3	-	34.75/.9300	30.59/.8476	29.33/.8112	28.93/.8671	34.30/.9496
CSNLN	x3	-	34.74/.9300	30.66/.8482	29.33/.8105	29.13/.8712	34.45/.9502
RFANet	x3	-	34.79/.9300	30.67/.8487	29.34/.8115	29.15/.8720	34.59/.9506
HAN	x3	-	34.75/.9299	30.67/.8483	29.32/.8110	29.10/.8705	34.48/.9500
EBPN	x3	-	-	-	-	-	-
IRN	x3	-	-	-	-	-	-
EDSR	x4	43.1M	32.46/.8968	28.80/.7876	27.71/.7420	26.64/.8033	31.02/.9148
RCAN	x4	15.6M	32.63/.9002	28.87/.7889	27.77/.7436	26.82/.8087	31.22/.9173
SAN	x4	15.7M	32.64/.9003	28.92/.7888	27.78/.7346	26.79/.8068	31.18/.9169
CSNLN	x4	3M	32.68/.9004	28.95/.7888	27.80/.7439	27.22/.8168	31.43/.9201
RFANet	x4	~11M	32.66/.9004	28.88/.7894	27.79/.7442	26.92/.8112	31.41/.9187
HAN	x4	-	32.64/.9002	28.90/.7890	27.80/.7442	26.85/.8094	31.42/.9177
EBPN	x4	~7M	32.79/.9032	29.01/.7903	27.85/.7464	27.03/.8114	31.53/.9198
IRN	x4	4.35M	36.19/.9451	32.67/.9015	31.64/.8826	31.41/.9157	-

对IRN，作者用高分辨率HR图像进行输入，而不是真实情况下的低分辨率LR图像，用后者输入时，性能会降低，具体见后续的“IRN性能提升检验”介绍。

---

论文简介和链接

论文笔记：

1. CVPR2020超分辨率方向论文整理笔记
2. ECCV2020超分辨率方向论文整理笔记
3. ICCV2019超分辨率方向论文整理笔记

• EDSR：Enhanced Deep Residual Networks for Single Image Super-Resolution，paper，code

  NTIRE2017超分冠军方案，使用增强的ResNet，去除Batch Norm。

• RCAN：Image Super-Resolution Using Very Deep Residual Channel Attention Networks，paper，code

  ECCV2018超分冠军方案，EDSR的改进，加入通道注意力。

• SAN：Second-Order Attention Network for Single Image Super-Resolution，paper，code

  CVPR2019，RCAN的改进，使用二阶注意力。

• CSNLN：Image Super-Resolution with Cross-Scale Non-Local Attention and Exhaustive Self-Exemplars Mining，paper，code

  CVPR2020，网络新基本块：跨尺度非局部注意力块和自样本挖掘机制。

• RFANet：Residual Feature Aggregation Network for Image Super-Resolution，paper

  CVPR2020，网络框架基于EDSR，其将EDSR中多个残差块加入块间的残差连接，然后加入空间注意力模块，性能超过了RCAN和SAN。

• HAN：Single Image Super-Resolution via a Holistic Attention Network，paper

  ECCV2020，全注意力机制：层、通道和位置的整体相互依赖关系进行建模。

• EBPN：Embedded Block Residual Network: A Recursive Restoration Model for Single-Image Super-Resolution，paper

  ICCV2020，两分支的嵌入残差模块：分别恢复低频和高频信息，并将前层难以恢复的信息传入更深的层进行恢复。

• IRN：Invertible Image Rescaling，paper，code

  ECCV2020，可逆神经网络，将HR经小波变换分解为低频分量和高频分量作为网络的输入，网络生成潜在分布和多个LR图像，然后以此再逆变换重建HR图像，性能提升极大。

另外几篇可参考的相关论文

CVPR2020

DRN：Closed-loop Matters: Dual Regression Networks for Single Image Super-Resolution，paper，code

​ 闭环问题：双监督方案。HR图像监督生成的SR图像，SR图像反过来受到LR图像监督。

UDVD：Unified Dynamic Convolutional Network for Super-Resolution with Variational Degradations，paper

​ 参考SRMD，训练数据重制作，LR图像包含多种降质；动态滤波器，多级损失。

ECCV2020

CDC：Component Divide-and-Conquer for Real-World Image Super-Resolution，paper，code

​ 注意力机制改进：注意力组件分而治之。构建三个分别与平面，边缘和角相关联的组件注意力块实现真实世界的图像超分辨率。

LatticeNet：LatticeNet: Towards Lightweight Image Super-resolution with Lattice Block，paper

​ 轻量级晶格网络：双支路特征提取基本块，反向级联融合策略。

Learning with Privileged Information for Efficient Image Super-Resolution，paper，code

​ 教师/学生网络：以HR图像信息指导学生网络的学习。

ICCV2020

KMSR：Kernel Modeling Super-Resolution on Real Low-Resolution Images，paper，code

​ 模糊核估计：用GAN估计HR的模糊核池，低质的HR同尺寸图像作为CNN网络的输入。

Wavelet Domain Style Transfer for an Effective Perception-distortion Tradeoff in Single Image Super-Resolution，paper

​ 小波变换：将低质高感知图像，高感知低质的图像分别进行小波变换，生成高质高感知的图像。