Artifact Reduction 和 Super Resolution的相关论文

Image Compression Improvement

• Learning Convolutional Networks for Content-weighted Image Compression

pdf

ImageCompression testcode

本文将加权重要性图和二进制量化应用在图像压缩系统中, 在图像压缩时引入important map来引导比特率分配，结果比jpeg、jpeg2000保留更多细节

• FARIP: Framework for Artifact Removal for Image Processing Using JPEG

本文优化了JPEG压缩方法

• Improved Lossy Image Compression with Priming and Spatially Adaptive Bit Rates for Recurrent Networks

Artifact Reduction based on CNN

• 2018-CVPR: BlockCNN: A Deep Network for Artifact Removal and Image Compression

ARmodel:输入24X24X3图像块，输出8X8X3图像块（3通道）。网络有一系列的卷积和残差块。残差块包含多个操作，包括卷积、批量归一化和leaky ReLU激活函数。其采用9个残差块，6百万个图像对做训练集。lr=10-3 use Adam with weightdecay of 10−4. 120, 000 iterations.和ARCNN做主观对比差异不大，无客观对比，无性能分析，无开源代码。

• 2017ICCV: Deep Generative Adversarial Compression Artifact Removal

利用生成对抗网络GAN做AR,效果优于ARCNN,能保留更多细节，无性能分析。两个网络，生成网络采用多层残差块，引入MSEloss+SSIMloss；判决网络采用多层卷积，输出2X2X1的判决结果，引入perceptualloss+AdversarialPatchloss；训练集16随机128×128 patches, with flipping and rotation data augmentation.learning rate of 10−4. 70, 000 iterations.无开源代码

• 2016arxiv: cas-cnn

提出了一种的12层深度卷积网络具有分层跳过的卷积层连接和定义multi-scale MSE loss。训练集：50k images of the 120 ×120 pixel。

• 2018arxiv: One-Two-One Networks for Compression Artifacts Reduction in Remote Sensing

应用于遥感图像的OTO网络：利用差分获得高频信息，利用求和来聚合低频信息。网络复杂度较高，其网络嵌套其他网络模型：densenet/resnet,共3路卷积网络相加/相减再处理。

CNN Network Optimization

• 2017TRANSACTIONS ON COMPUTATIONAL IMAGING: Loss Functions for Image Restoration With Neural Networks

作者提出对于图像恢复算法的神经网络中，提出Multiscale SSIM loss,建议使用l1 loss+MSSSIMloss 能获得更好的恢复效果，能够有效去除块效应，尤其在天空等颜色变化较小的区域。

Video Restoration

• 2018TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY: Understanding and Removal of False Contour in HEVC Compressed Images

应用传统方法去伪影

Image Restoration and SR

• Underwater Image Restoration Based on Image Blurriness and Light Absorption

• 2017TIP：Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising

• code

• code-tensorflow

作者介绍该网络可以有效去噪、去压缩失真、超分辨率。该网络利用20层conv-BN-RELU层输出残差，计算残差后与原图相加得到恢复图像.
128×8,000 image patch (the size is 50 × 50) pairs for training。 denoise、SR、AR效果均比经典算法psnr明显提升；计算时间(GPU):512x512-0.06S;1024X1024-0.235S

• 2018icip：DUAL-DOMAIN MULTI-SCALE CONVOLUTIONAL NEURAL NETWORK FOR COMPRESSION ARTIFACTS REMOVAL

• Project Page

主观效果比arcnn/DNCNN/DDCN/TNRD要好，但网络较复杂，未有时间复杂度的对比。

• 2017cvpr: Beyond Deep Residual Learning for Image Restoration: Persistent Homology-Guided Manifold Simplification

应用于图像去噪/图像超分辨率

• 从SRCNN到EDSR，总结深度学习端到端超分辨率方法发展历程

• DRRN: Image Super-Resolution via Deep Recursive Residual Network, CVPR2017

• code

选用的是1个递归块和25个残差单元，深度为52层的网络结构。是通过对之前已有的ResNet等结构进行调整，采取更深的网络结构得到结果的提升。

• LapSRN: Deep Laplacian Pyramid Networks for Fast and Accurate Super-Resolution, CVPR2017

• code

• SRGAN: Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network, CVPR2017

Code:

github(tensorflow): https://github.com/zsdonghao/SRGAN
github(tensorflow): https://github.com/buriburisuri/SRGAN
github(torch): https://github.com/junhocho/SRGAN
github(caffe): https://github.com/ShenghaiRong/caffe_srgan
github(tensorflow): https://github.com/brade31919/SRGAN-tensorflow

SRGAN+WGAN

• EDSR: Enhanced Deep Residual Networks for Single Image Super-Resolution, CVPR2017