- Restoration
- Image Enhancement
- Low Light
- Denoising
- Debluring
- Dehazing
- Super Resolution
- Image Synthesis
- General
- HDR
- Using Raw
- ISP
- Dataset
- Useful Resources
[Paper] (CVPR 2018) Learning Dual Convolutional Neural Networks for Low-Level Vision
[Author] Jinshan Pan, Sifei Liu, Deqing Sun, Jiawei Zhang, Yang Liu, Jimmy Ren, Zechao Li, Jinhui Tang, Huchuan Lu, Yu-Wing Tai, Ming-Hsuan Yang
[Project] [Unofficial-TF-Code]
粗读, 设计了一双分支网络, 一个学习detail, 一个学习structure, 针对任务对两个分支也分别进行监督训练
[Paper] (CVPR 2018) Deep Image Prior
[Author] Dmitry Ulyanov, Andrea Vedald, Victor Lempitsky
[Project]
- 一篇有趣的论文, 提出深度卷积网络在图像生成和恢复任务中表现好的原因, 可能并不是因为其从大量图像中学习到了某种先验, 其实随机初始化的网络足以从输入中抓取大量的low-level图像先验信息. 在通过迭代的方式从图像中学习先验的过程中, 那些自然的, 有规律的内容较容易提取,会先被学习出来, 因此就达到了去噪或其它restoration的目的.
- 粗读, 实用性有待验证, 有时间可以好好研究一下.
[Paper] (CVPR 2019) Dual Residual Networks Leveraging the Potential of Paired Operations for Image Restoration
[Author] Xing Liu, Masanori Suganuma, Zhun Sun, Takayuki Okatani
[Code]
- 文章提出, 许多图像复原任务都由一些成对的模块组成, 比如去噪里的大kernel和小kernel, 超分里的下采样和上采样. 本文在residual connection的基础上, 进一步给每个模块内部的操作直接加入residual connection, 增加了组合数.
- 在去噪, 去模糊, 去雾等任务中都取得了不错的效果.
[Paper] (CVPR 2019) Attention-based Adaptive Selection of Operations for Image Restoration in the Presence of Unknown Combined Distortions
[Author] Masanori Suganuma, Xing Liu, Takayuki Okatani
[Code]
- 提出用一个基于attention的操作加权网络, 用来处理不同种类的degradation.
- 性能一般, 不太容易收敛, 思路值得借鉴.
[Paper] (ICCV 2019) CFSNet: Toward a Controllable Feature Space for Image Restoration
[Author] Wei Wang, Ruiming Guo, Yapeng Tian, Wenming Yang
[Pytorch-Code]
粗读, 用一个手动输入的参数控制两个分支的权重, 一个分支负责low distortion修复, 另一个分支负责high visual quality. 两个分支通过使用不同loss (L1, L2 v.s. vgg, GAN loss) 训练来得到. 文章的效果和实用性有待检验, 思路可借鉴.
[Paper] (WACV 2019) Gated Context Aggregation Network for Image Dehazing and Deraining
[Author] Dongdong Chen, Mingming He, Qingnan Fan
[Code]
[Paper] (BMVC 2014) Reproduction Angular Error: An Improved Performance Metric for Illuminant Estimation
[Author] Graham Finlayson, Roshanak Zakizadeh
- 提出了一个用于评估illuminant estimation性能的准则, 该准则与光源的色温无关. 大致浏览, 一些原理没看懂.
- 后面Google在此基础上做了改进, 作为loss去训练低光照时AWB模型.
[Paper] (CVPR 2018) Learnign a Discriminative Prior for Blind Image Deblurring
[Author] Lerenhan Li, Jinshan Pan, Wei-Sheng Lai, Changxin Gao, Nong Sang Ming-Hsuan Yang
[Project]
用CNN学习一个deblur用的prior, 用来提供输入图像是否模糊的先验知识, 把该prior加入目标函数, 之后用迭代的方法求解优化函数
[Paper] (ICCV 2017) DSLR-Quality Photos on Mobile Devices with Deep Convolutional Networks
[Author] Andrey Ignatov, Nikolay Kobyshev, Kenneth Vanhoey, Radu Timofte , Luc Van Gool
[Code]
- 从变换的角度出发, 学习一个从低质量图像到高质量图片的变换函数
- 变换部分采用残差快结构的CNN，定义了4个loss (color, texture, content, variance). color loss是图像进行高斯模糊后的均方差, texture loss是adversarial loss, content loss是perceptual loss, variance loss是图像梯度的模.
- 提出了用于图像质量增强的数据集DPED, 包括iPhone, BlackBerry和Sony三种手机与Canon单反相机的图相对.
[Paper] (arXiv 1707) Aesthetic-Driven Image Enhancement by Adversarial Learning
[Author] Yubin Deng, Chen Change Loy, Xiaoou Tang
- weakly supervised方法, 学习crop和色彩变换参数, 增强aesthetic quality
[Paper] (CVPRW 2018) WESPE: Weakly Supervised Photo Enhancer for Digital Cameras
[Author] Andrey Ignatov, Nikolay Kobyshev, Radu Timofte , Kenneth Vanhoey, Luc Van Gool
[Project]
- 弱监督, 训练时无需成对的低质量图像和高质量图像. 用两个adversarial losses (color和texture)保证将低质量图像变换到高质量图像所在的域
- 定义一content loss保证增强后的图像与输入图像的content consistency. 注意此处是将增强后的图像backward map到输入空间, 在输入空间定义的perceptual loss
- 定义一total variation (TV)保证输出的平滑
- 本文的思路及loss的设计来自DPED
[Paper] (ECCVW 2018) Range Scaling Global U-Net for Perceptual Image Enhancement on Mobile Devices
[Author] Jie Huang, Pengfei Zhu, Mingrui Geng, Jiewen Ran, Xingguang Zhou, Chen Xing, Pengfei Wan, Xiangyang Ji
[TF-Code]
UNet + Global Pooling feature + 输入输出feature间的elementwise scaling
[Paper] (TIP 2018) Learning a Deep Single Image Contrast Enhancer from Multi-Exposure Images
[Author] Jianrui Cai, Shuhang Gu, Lei Zhang
[Caffe-Code]
[Paper] (FG 2018) GLADNet: Low-Light Enhancement Network with Global Awareness
[Author] Wenjing Wang,Chen Wei, Wenhan Yang, Jiaying Liu
[Page] [TF-Code]
encoder-decoder + refine结构的网络
[Paper] (BMVC 2018 Oral) Deep Retinex Decomposition for Low-Light Enhancement
[Author] Chen Wei, Wenjing Wang, Wenhan Yang, Jiaying Liu
[Page] [TF-Code]
基于retinex理论设计的网络, 后续一些工作基于这个思路展开, 但本文的效果一般
[Paper] (BMVC 2018) MBLLEN: Low-light Image/Video Enhancement Using CNNs
[Author] Feifan Lv, Feng Lu, Jianhua Wu, Chongsoon Lim
[Page] [TF-Code]
多分支亮度增强网络
[Paper] (CVPR 2018 Spotlight) Deep Photo Enhancer: Unpaired Learning for Image Enhancement from Photographs with GANs
[Author] Yu-Sheng Chen, Yu-Ching Wang, Man-Hsin Kao, Yung-Yu Chuang
[TF-Code] [TF-Code2]
UNet + cycGAN, 无需paired样本的图像增强方法, 可以参考, 只是代码有一点点乱
[Paper] (NIPS 2018) DeepExposure: Learning to Expose Photos with Asynchronously Reinforced Adversarial Learning
[Author] Runsheng Yu, Wenyu Liu, Yasen Zhang, Zhi Qu, Deli Zhao, Bo Zhang
[Paper] (MM 2019) Kindling the Darkness: A Practical Low light Image Enhancer
[Author] Yonghua Zhang, Jiawan Zhang, Xiaojie Guo
[TF-Code-KinD] [TF-Code-KinD++]
- 采用类似Retinex的结构, 两个分支分别预测亮度分量和反射分量. 网络结构和loss可以参考.
- 提出了一个小型的亮度adjustment net, 可以输入一个ratio, 控制增强程度, 比较有趣.
[Paper] (CVPR 2019) Underexposed Photo Enhancement Using Deep Illumination Estimation
[Author] Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, Jiaya Jia
[TF-code]
同样基于Retinex理论, 但网络只预测illumination map, 使用了reconstruction, color和smooth loss. 整个工程都建立在HDRNet的基础上. 用联合上采样的思路做tone mapping的思路感觉可以挖掘一下.
[Paper] (arXiv 1906) EnlightenGAN: Deep Light Enhancement without Paired Supervision
[Author] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, Zhangyang Wang
[Pytorch-code]
基于GAN的非监督亮度增强方法, 效果不错
[Paper] (arXiv 1906) Low-light Image Enhancement Algorithm Based on Retinex and Generative Adversarial Network
[Author] Yangming Shi, Xiaopo Wu, Ming Zhu
RetinexNet+GAN
[Paper] (arXiv 1908) Attention Guided Low-light Image Enhancement with a Large Scale Low-light Simulation Dataset
[Author] Feifan Lv, Yu Li, Feng Lu
[Page]
- 同时做tone mapping和去噪, 分为亮度attention map预测, noise map预测, 多尺度增强模块和refine模块四部分, 网络结构和loss可以参考
- 提出了一个生成低光照加噪声数据的流程.
[Paper] (arXiv 1911) Color-wise Attention Network for Low-light Image Enhancement
[Author] Yousef Atoum, Mao Ye, Liu Ren, Ying Tai, Xiaoming Liu
亮度和颜色通道分两只分别增强的方案, 其中color和point的attention部分没看懂
[Paper] (CVPR 2020) Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement
[Author] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, Runmin Cong
[Page] [Code]
- 一篇挺有趣的论文, 把tone mapping看成pixel-wise的曲线预测问题, 设计了一个小型曲线估计网络, 并提出了几个无监督loss, 得到了不错的结果
- 一些局限性: 提出的一系列约束loss对于增强部分区域可能不太适用, 比如对夜景图片增强前景的同时保持夜空是暗的
[Paper] (arXiv 2003) Learning to Correct Overexposed and Underexposed Photos
[Author] Mahmoud Afifi, Konstantinos G. Derpanis, Björn Ommer, Michael S. Brown
[Code]
粗读, coarse-to-fine增强的策略, 并在每个level加入相应的拉普拉斯金字塔层作为细节信息. 使用L1和GAN loss. 效果不错.
[Paper] (CVPR 2017) Learning Deep CNN Denoiser Prior for Image Restoration
[Author] Kai Zhang, Wangmeng Zuo, Shuhang Gu, Lei Zhang
[Matlab-Code]
大致浏览, 提出了一个结构简单的CNN去噪器, 可以为基于模型的优化方法提供有效的prior, 还可以用于求解其它图像恢复的逆问题
[Paper] (ICCV 2019 Oral) Real image denoising with feature attention
[Author] Saeed Anwar, Nick Barnes
[Code]
- 提出了一个端到端的去噪网络, 基于channel attention和skip connection. 在真是图像上测试效果不错, 速度一般.
- 作为一篇Oral来说感觉创新点和理论论述都一般, 也没有解释为什么提出的网络对真是图像去噪效果好.
- 如果需要, 参考网络流程图和代码即可.
[Paper] (CVPR 2019) Unprocessing Images for Learned Raw Denoising
[Author] Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, Jonathan T. Barron
[Code]
- 提出了一个通过unprocess ISP流程而生成更真实去噪样本的框架, 可以用任意图像生成真实的训练样本, 以提高模型性能.
- 对于sRGB图像, 根据ISP流程, 将其逐步逆运算位raw image, 在此基础上加的噪声更符合真实噪声.
- 推断时, 要先把sRGB转换为raw image, 再经过网络处理, 最后再进行正向的ISP恢复为sRGB.
- ISP流程的推断对每个品牌型号的相机都有所不同, 模拟其过程感觉还是有难度的.
[Paper] (CVPR 2019) Toward Convolutional Blind Denoising of Real Photographs
[Author] Shi Guo, Zifei Yan, Kai Zhang, Wangmeng Zuo, Lei Zhang, Jonathan T. Barron
[Code]
- 大致浏览. 采用一个FCN估计噪声level, 噪声level map与输入concat然后输入一类似U-Net的网络去噪.
- 可以学习其网络和训练细节.
[Paper] (CVPR 2017 Spotlight) Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring
[Author] Seungjun Nah, Tae Hyun Kim, Kyoung Mu Lee
[Code]
- 提出了GOPRO单张图像deblur数据集
- 提出了一个多尺度输入的去噪网络
[Paper] (ICASSP 2018) A Deep Encoder-Decoder Network For Joint Deblurring and Super-Resolution
[Author] Xinyi Zhang, Fei Wang, Hang Dong, Yu Guo
[Project]
大致浏览, 一个端到端的同时deblur和超分网络
[Paper] (BMVC 2018) Gated Fusion Network for Joint Image Deblurring and Super-Resolution
[Author] Xinyi Zhang, Hang Dong, Zhe Hu, Wei-Sheng Lai, Fei Wang, Ming-Hsuan Yang
[Project] [Pytorch-Code]
- 提出了一个同时做deblur和超分的网络. 网络有两个分支, 一个encoder-decoder结构做deblur, 一个不降分辨率做SR, 用一个几层卷积组成的gate模块选择特征.
- 思路简单, 可以尝试.
[Paper] (CVPR 2018) DeblurGAN: Blind Motion Deblurring Using Conditional Adversarial Networks
[Author] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, Jiří Matas
[Pytorch-Code] [Unofficial-TF-Code1] [Unofficial-TF-Code2]
- 用GAN做deblur的一篇典型文章, 效果不错.
- 生成网络结构简单, 采用残差形式.
- 提出了生成blur数据的方法, 可以参考一下.
[Paper] (ICCV 2019) DeblurGAN-v2: Deblurring (Orders-of-Magnitude) Faster and Better
[Author] Orest Kupyn, Tetiana Martyniuk, Junru Wu, Zhangyang Wang
[Pytorch-Code]
DeblurGAN基础上的改进, 把生成网络换成了FPN, 设计了新的loss, 效果更快更好了
[Paper] (WACV 2019) Gyroscope-Aided Motion Deblurring with Deep Network
[Author] Janne Mustaniemi, Juho Kannala, Simo Särkkä, Jiri Matas, Janne Heikkilä
结合陀螺仪作为先验deblur. 从陀螺仪和图像拍摄信息生成训练集的方法可以参考.
[Paper] (CVPR 2019) Douglas-Rachford Networks: Learning Both the Image Prior and Data Fidelity Terms for Blind Image Deconvolution
[Author] Raied Aljadaany, Dipan K. Pal, Marios Savvides
- 基于Douglas-Rachford迭代优化求解blind deconvolution的思路(不懂), 提出了一个由简单conv和连接操作组成的Dr Block, 将其嵌入普通卷积网络中, 用L2和GAN loss训练, 取得了不错的效果.
- 网络细节没看, 可以借鉴其模块设计
[Paper] (CVPR 2019) Deep Stacked Multi-patch Hierarchical Network for Image Deblurring
[Author] Hongguang Zhang, Yuchao Dai, Hongdong Li, Piotr Koniusz
[Pytorch-Code]
从spatial pyramid matching的角度出发, 提出了一个分patch的逐层融合处理的网络, 参数少速度快. 但个人仍不理解这种分patch的做法对CNN来说到底有什么意义.
[Paper] (ICCV 2019) Human-Aware Motion Deblurring
[Author] Ziyi Shen, Wenguan Wang, Xiankai Lu, Jianbin Shen, Haibin Ling, Tingfa Xu, Ling Shao
[Project] [HIDE Dataset]
- 提出了HIDE数据集, 主要关注对人体的deblur
- 提出了一个多分支deblur网络, 根据human-aware子网络预测前背景生成weight map, 将多分枝信息融合处理后输出
[Paper] (CVPR 2018) Single Image Dehazing via Multi-Scale Convolutional Neural Networks
[Author] Wenqi Ren, Si Liu, Hua Zhang, Jinshan Pan, Xiaochun Cao, Ming-Hsuan Yang
[Project] [Matlab-Code] [Unofficial-TF-Code]
大致浏览, 一个多尺度去雾网络, coarse尺度预测transmission map, fine尺度预测去雾图像, 用深度图生成transmmision map训练
[Paper] (CVPR 2019) Blind Super-Resolution with Iterative Kernel Correction
[Author] Jinjin Gu, Hannan Lu, Wangmeng Zuo, Chao Dong
[Project]
- 粗读, 提出一个基于深度学习的交替预测blur kernel和预测超分结果的模型, 对给定的blur有很好的效果
- 文中提出的预测blur kernel并用其辅助超分的思路很有意思, 但对真实图像而言无法获得真实的blur kernel用于训练, 另外论文似乎假设一张图像只有一种blur kernel, 感觉不太合理
[Paper] (CVPR 2019) Camera Lens Super-Resolution
[Author] Chang Chen, Zhiwei Xiong, Xinmei Tian, Zheng-Jun Zha, Feng Wu
[Code & Data]
文章认为普通的插值退化不能模拟由于焦距-FOV变化带来的退化 (其实这是一个无论从分析上还是工程中都很明显的事实...). 最重要的贡献是提出了一个真实DSLR和手机的数据集, 但是在生成单反数据集时, 貌似没有考虑焦距变化带来的景深变化.
[Paper] (ICCV 2017) Dynamic-Net: Tuning the Objective Without Re-training for Synthesis Tasks
[Author] Qifeng Chen, Jia Xu, Vladlen Koltun
[Project] [PyTorch-Code]
先以Objective 0训练主干网络, 之后固定主干网络以Objective 1训练tuning block. 测试时手动指定插值系数, 达到在O0和O1之间的输出效果. 论文思路和实现都很简单, 分析论述方式值得学习
[Paper] (ICCV 2017) Fast Image Processing with Fully-Convolutional Networks
[Author] Qifeng Chen, Jia Xu, Vladlen Koltun
[Project]
较早用CNN做图像滤波增强的paper之一, 使用了dilation conv提取全局信息.
[Paper] (CVPR 2018) Fast End-to-End Trainable Guided Filter
[Author] Huikai Wu, Shuai Zheng, Junge Zhang, Kaiqi Huang
[Code]
- 可训练的引导滤波, 用于联合上采样. 可用于各种像素级的增强任务中.
- 同时提供了用TensorFlow实现的原始guided filter, 赞！
[Paper] (CVPR 2018) The Perception-Distortion Tradeoff
[Authors] Yochai Blau, Tomer Michaeli
- 大致浏览, 提出在image restoration中, perception和distortion存在tradeoff. 对不同的loss这种tradedoff的严重程度不同, 如perceptual loss与MSE loss相比能在perception和distortion直接取得更好的平衡.
- 很多理论都还没看, 日后如果研究这一方向, 可以仔细读一下.
[Paper] (ECCV 2018) Decouple Learning for Parameterized Image Operators
[Author] Qingnan Fan, Dongdong Chen, Lu Yuan, Gang Hua, Nenghai Yu, Baoquan Chen
[PyTorch-Code]
粗读, 貌似是给不同任务设定一个parameter, 用网络以parameter为输入预测每层的weight, 这个weight作为instance norm的weight对每层做归一化.
[Paper] (ECCV 2018 Oral) The Contextual Loss for Image Transformation with Non-Aligned Data
[Authors] Roey Mechrez, Itamar Talmi, Firas Shama, Lihi Zelnik-Manor
[Project] [Code]
提出了一个处理非对齐数据的loss, 利用特征(用VGG19获得)的距离定义两像素特征点的相似度, 并在此基础上定义loss, 以解决输入和真值在空间上不对齐的问题.
[Paper] (CVPR 2019) Spatially Variant Linear Representation Models for Joint Filtering
[Author] Jinshan Pan, Jiangxin Dong, Jimmy S. Ren, Liang Lin, Jinhui Tang, Ming-Hsuan Yang
[Project]
大致浏览, 用CNN预测guided filter中的系数A和b.
[Paper] (ICCV 2019) Self-Guided Network for Fast Image Denoising
[Author] Shuhang Gu, Yawei Li, Luc Van Gool, Radu Timofte
[Pytorch-Code]
[Paper] (ICCV 2019) Fast Image Restoration with Multi-bin Trainable Linear Units
[Author] Shuhang Gu, Wen Li, Luc Van Gool, Radu Timofte
[Pytorch-Code]
[Paper] (ECCV 2010) Guided Image Filtering
[Author] Kaiming He, Jian Sun, Xiaoou Tang
[Project] [TF/Pytorch-Code]
大名鼎鼎的引导滤波, 可用在去噪, 融合, 联合上采样, matting, 图像增强等多种任务中. 速度快, 效果好.
[Paper] (SIGGRAPH 2011) Local Laplacian Filters: Edge-aware Image Processing with a Laplacian Pyramid
[Author] Sylvain Paris, Samuel W. Hasinoff, Jan Kautz
[Project] [Code1] [Code2]
用拉普拉斯金字塔做图像增强, tone mapping等.
[Paper] (ICCV 2017) Misalignment-Robust Joint Filter for Cross-Modal Image Pairs
[Author] Takashi Shibata, Masayuki Tanaka, Masatoshi Okutomi
- 提出一种适用于不对齐多模数据的联合滤波方法, 可结合引导滤波等优良滤波算法, 在非对齐不同源数据上达到很好的滤波效果.
- 算法的思路其实就是计算cost volume并对其进行加权求和. 其最好版本的大体思路为: 将引导图上下左右位移组成k个移位引导图, 1.计算target和k个引导图的距离(NCC等)组成cost volume. 2.从cost volume计算weight volume, 并通过最小化能量函数的方法对其进行优化. 3.用k个移位引导图分别对target进行滤波.4.用weight volume对k个滤波输出进行加权平均, 生成最后的输出.
- 从paper中看, 该方法对非对齐的多模数据滤波效果不错, 可以在设计DL方案时作为参考.
- 算法的局限: 1.weight volume优化的步骤过于耗时; 2. cost volume的准确性仍依赖于距离的计算准则, 现有的例如NCC等策略也不能完美解决多模数据的相似性度量问题.
[Paper] (Siggraph Asia 2017) HDR image reconstruction from a single exposure using deep CNNs
[Author] Gabriel Eilertsen, Joel Kronander, Gyorgy Denes, Rafał K. Mantiuk, Jonas Unger
[Page] [TF-Code]
粗读, 较早将CNN用于HDR的一篇paper, 网络结构比较老, 但提出了两个可能有趣的点: 1) 将输入转换到log域训练更符合人眼视觉特性; 2) 高光部分采用输入和输出线性加权的方式, 修复过曝光, 并避免形成带状伪影.
[Paper] (Siggraph 2017) Deep Bilateral Learning for Real-Time Image Enhancement
[Author] Michaël Gharbi, Jiawen Chen, Jonathan T. Barron, Samuel W. Hasinoff, Frédo Durand
[Project]
- 提出了一个实时图像增强网络, 速度快, 效果好.
- 网络分为两个分支, 低分辨率分支提取特征, 学习每个像素的色彩映射参数; 高分辨率分支负责提取和保留细节信息. low res分支学到的映射参数通过类似于双线性差值的过程上采样到high res, 最后对high res图像做色彩映射并输出.
- 学习映射参数部分, 采用bilateral grid的思路. 第三个维度被解释成8*12的网格, 意思是对8个灰度level做不同的色彩映射. 处理时选择哪个level的参数, 由high res分支生成的引导图决定.
[Paper] (Siggraph Asia 2017) Deep Reverse Tone Mapping
[Author] Yuki Endo, Yoshihiro Kanamori, Jun Mitani
[Project] [Unofficial-Pytorch-Code]
[Paper] (EG 2018) ExpandNet: A Deep Convolutional Neural Network for High Dynamic Range Expansion from Low Dynamic Range Content
[Author] Demetris Marnerides, Thomas Bashford-Rogers, Jonathan Hatchett, Kurt Debattista
[Pytorch-Code] [Unofficial-TF-Code]
[Paper] (CVPR 2020) Single-Image HDR Reconstruction by Learning to Reverse the Camera Pipeline
[Author] Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, Jia-Bin Huang
[Project] [TF-Code]
用几个CNN模拟ISP中HDR到LDR的映射过程, 完成HDR. 关键是定义每个部分的训练目标和数据, 这部分没细看.
[Paper] (CVPR 2019) Towards Real Scene Super-Resolution with Raw Images
[Author] Xiangyu Xu, Yongrui Ma, Wenxiu Sun
[Project]
大致浏览, 利用Raw做细节恢复, 用RGB做Color校正.
[Paper] (arXiv 1905) Trinity of Pixel Enhancement: a Joint Solution for Demosaicking, Denoising and Super-Resolution
[Author] Guocheng Qian, Jinjin Gu, Jimmy Ren, Chao Dong, Furong Zhao, Juan Lin
[Pytorch-Code]
- 使用具有pixel shift技术的相机收集了一可以做demoasic的数据集, 避免了用普通RGB数据做真值时内置demoasic过程带来的误差
- 提出了一端到端的demosaic, 去噪和超分的网络, 采用residual + dense block的形式, 没什么特别的
[Paper] (TIP 2018) DeepISP: Learning End-to-End Image Processing Pipeline
[Author] Eli Schwartz, Raja Giryes, Alex M. Bronstein
大致浏览, 一个end-to-end的网络, 分为保持分辨率的low level部分和逐层下采样的high level部分. 使用了conv+relu, conv+tanh, 直连三个分支并行的设计, 比较少见
[Paper] (CVPR 2018) Learning to See in the Dark
[Author] Chen Chen, Qifeng Chen, Jia Xu, Vladlen Koltun
[Project] [TF-Code]
- 提出了SID数据集, 包括RGB和Raw数据
- 提出了一个end-to-end的isp网络, 以RAW和增益信息为输入, 输入RGB图像, 代替传统ISP流程
[Paper] (arXiv 2002) Replacing Mobile Camera ISP with a Single Deep Learning Model
[Author] Andrey Ignatov, Luc Van Gool, Radu Timofte
[Code]
- 提出了一个端到端的深度学习网络, 用以代替现有的ISP处理流程.
- 提出了一个华为P20 RAW 和Canon 5D的RAW-RGB图像对, 用以训练ISP模型.
- 提出的算法与自带的ISP流程相比, 色彩上有一定提升, 但没有明显优势, 且存在晕影. 另外速度也是个问题. 因此对于用一个DL模型代替ISP流程的方案可行性还是有待确认.
PolyU
Renoir
CC
SID
kodak_color
NoiseClinicImages
HIDE motion deblur
GOPRO motion deblur
https://paperswithcode.com/task/image-denoising?page=2
https://github.com/wenbihan/reproducible-image-denoising-state-of-the-art
https://github.com/subeeshvasu/Awesome-Deblurring
https://github.com/BBuf/Image-processing-algorithm
https://github.com/Ir1d/lowLevelVision/tree/3ff10054beb7b83f74a1ca11d84562e3ea90d273
https://github.com/Elin24/Awesome-Low-Light-Enhancement
[ISP介绍blog] https://blog.csdn.net/qq_42261630/article/details/102918149