GRDN:Grouped Residual Dense Network for Real Image Denoising and GAN-based Real-world Noise Modeling

---

title: paper reading001| GRDN——Ntire real image denoising champion
date: 2019-07-29 10:47:44
categories: “论文阅读”
mathjax: true

GRDN:Grouped Residual Dense Network for Real Image Denoising and GAN-based Real-world Noise Modeling

arXiv

2019 CVPR Workshop

PyTorch

@InProceedings{Kim_2019_CVPR_Workshops,
author = {Kim, Dong-Wook and Ryun Chung, Jae and Jung, Seung-Won},
title = {GRDN:Grouped Residual Dense Network for Real Image Denoising and GAN-Based Real-World Noise Modeling},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2019}
}

关键词

作者的摘要和介绍部分主要是论述，真实噪声图像和其对应的GT是十分难获得的，仿真的合成数据和真实数据有很大的GAP，现阶段面向真实图像去噪的方法主要有两大类

• 基于一个更好的真实图像噪声数据模型statistical model of real-world(其实就是仿真数据)

优点：生成成对训练数据容易

缺点：是否真实图像噪声可以被模型化为一个统计模型依然存在争议

列举的五篇文章，都是使用仿真数据的，前面三篇比较出名，后面两篇早期TIP的以后可以读一下，应该大概都是说高斯泊松噪声模型、信号相关噪声和真实图像的噪声更为接近。

Unprocessing

Cbdnet

Neural nearest neighbors networks

Practical Poissonian-Gaussian noise modeling and fitting for single-image raw-data(TIP 2008)

Practical signal dependent noise parameter estimation from a single noisy image (TIP 2014)

• 真实噪声数据

From real-world noisy images, nearly noise-free ground-truth images can be obtained by inverting an image acquisition procedure

也同样列举五篇文章，基本就是五个数据集合

SIDD

RENOIR

DND

Learning to see in the dark(CVPR 2018)

arXiv

DeepISP: Toward learning an end-to-end image processing pipeline(TIP 2019)

arXiv

However, the amount of provided images may not be enough for training a large network and without a sufficient knowhow it is difficult to generate ground-truth images from real-world noisy images.

We thus adopt the second approach but applied our own generative adversarial network (GAN)-based data augmentation technique to obtain a larger dataset.

这里应该本文的第一个创新点：使用GAN方法进行数据增强

网络结构

Figure 1: The proposed network architecture: GRDN.

RDN - GRDN

convolutional block attention module (CBAM)

ECCV 2018 arXiv

rdn and grdn