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。


你可能感兴趣的:(医学图像超分辨率,图像高分辨率,深度学习,图像处理)