# Deep-Restore-PyTorch
Deep CNN for learning image restoration without clean data! (Noise2Noise)
## Introduction
Image restoration is task in which we have a noisy input image and we desire to get a noise free output image. Several techniques have been proposed for this task. One is using the Light Transport Simulation algorithm, which traces the path of millions of light rays. The disadvantage of this technique is the rendering time. It may take upto hours to render a single scene.</br>
Recently Convolutional Neural Networks have been proposed to solve this problem, which they do and also achieve state of the art results. They can learn the concept of noise, hence when provided with unseen noisy image, it can generate a noise free image easily in very less time. We can simply train a CNN for this task by providing noisy image as input and clean image as target and minimize a loss function using gradient descent.
![](http://zinggadget.com/wp-content/uploads/2018/07/Noise2Noise-Nvidia-Kecerdasan-Buatan.jpg)
## The Problem
It works. Works pretty well indeed. But there are cases where it is not just expensive but impossible to create clean images for our training data. Low light photography, Astronomical Imaging or Magnetic Resonance Imaging (MRI) are few of such cases. Neural Network based techniques cannot be easily used for this case.
## The Solution
Here comes a novel technique to address this issue. This method uses only noisy images to train the neural network to produce clean image as output. No clean images are required whatsoever for this technique.
## A Few Samples
### Multiplicative Bernoulli Noise Removal
![](imgs/index8.png) ![](imgs/index7.png)
### Gaussian Noise Removal
![](imgs/index4.png) ![](imgs/index3.png)
### Corrupt Text Removal
![](imgs/index2.png) ![](imgs/index.png)</br>
### Poisson Noise Removal
![](imgs/index6.png) ![](imgs/index5.png)
#### Note: I am not yet sure about how poisson noise needs to be implemented because it's not additive and is signal dependent while the paper simply mentions using a constant lambda.</br>
#### The minor artifacts seen are due to the following
1. Only 291 images were used for training.
2. The images were random cropped to 64 x 64 for quick training.</br>
Progressive resizing could be incorporated along with more images to get high resolution results.
## References:
[Noise2Noise Paper](https://arxiv.org/pdf/1803.04189.pdf)
没有合适的资源?快使用搜索试试~ 我知道了~
无需干净数据即可学习图像恢复的深度CNN!_Jupyter Notebook_Python_下载.zip
共414个文件
jpg:300个
bmp:91个
png:8个
1.该资源内容由用户上传,如若侵权请联系客服进行举报
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
版权申诉
0 下载量 158 浏览量
2023-04-28
13:37:46
上传
评论
收藏 38.08MB ZIP 举报
温馨提示
无需干净数据即可学习图像恢复的深度CNN!_Jupyter Notebook_Python_下载.zip
资源推荐
资源详情
资源评论
收起资源包目录
无需干净数据即可学习图像恢复的深度CNN!_Jupyter Notebook_Python_下载.zip (414个子文件)
000tt2.bmp 448KB
000tt3.bmp 445KB
000tt9.bmp 427KB
000tt4.bmp 417KB
000tt25.bmp 405KB
000tt1.bmp 398KB
000tt5.bmp 397KB
000tt6.bmp 383KB
000tt21.bmp 374KB
000tt8.bmp 371KB
000tt26.bmp 350KB
000t46.bmp 347KB
000tt12.bmp 335KB
000tt24.bmp 331KB
000t37.bmp 305KB
000tt13.bmp 286KB
000tt27.bmp 277KB
000tt23.bmp 265KB
000tt20.bmp 260KB
000t50.bmp 254KB
000t51.bmp 246KB
000t49.bmp 237KB
000tt10.bmp 236KB
000tt17.bmp 233KB
000t53.bmp 232KB
000t66.bmp 229KB
000tt15.bmp 211KB
000t59.bmp 209KB
000t52.bmp 202KB
000t42.bmp 198KB
000t40.bmp 195KB
000t38.bmp 195KB
000t39.bmp 193KB
000tt19.bmp 186KB
000tt16.bmp 172KB
000t55.bmp 169KB
000tt22.bmp 159KB
000t43.bmp 153KB
000tt14.bmp 150KB
000t54.bmp 150KB
000t63.bmp 145KB
000t47.bmp 140KB
000tt7.bmp 140KB
000t60.bmp 138KB
000t48.bmp 132KB
000t58.bmp 130KB
000t4.bmp 126KB
000t15.bmp 124KB
000t27.bmp 123KB
000t35.bmp 122KB
000t25.bmp 122KB
000t45.bmp 117KB
000t11.bmp 116KB
000t62.bmp 115KB
000t64.bmp 115KB
000t44.bmp 113KB
000t36.bmp 111KB
000t65.bmp 111KB
000t57.bmp 109KB
000t61.bmp 109KB
000t1.bmp 102KB
000t8.bmp 100KB
000t31.bmp 99KB
000t13.bmp 99KB
000t56.bmp 96KB
000t19.bmp 94KB
000tt18.bmp 94KB
000t28.bmp 90KB
000t2.bmp 90KB
000t6.bmp 88KB
000t3.bmp 87KB
000t29.bmp 85KB
000t32.bmp 85KB
000t7.bmp 82KB
000t16.bmp 82KB
000t33.bmp 80KB
000t14.bmp 75KB
000t5.bmp 69KB
000t17.bmp 68KB
000t34.bmp 65KB
000t30.bmp 60KB
000t22.bmp 58KB
000t26.bmp 57KB
000t18.bmp 54KB
000t10.bmp 54KB
000t23.bmp 46KB
000t21.bmp 40KB
000t12.bmp 40KB
000t9.bmp 37KB
000t24.bmp 36KB
000t20.bmp 18KB
.gitignore 57B
test_dataset.ipynb 637KB
test_dataset-checkpoint.ipynb 325KB
test_train.ipynb 216KB
test_train-checkpoint.ipynb 193KB
291000.jpg 123KB
365073.jpg 121KB
175032.jpg 118KB
147062.jpg 115KB
共 414 条
- 1
- 2
- 3
- 4
- 5
资源评论
快撑死的鱼
- 粉丝: 1w+
- 资源: 9154
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功