cszn / ircnn Goto Github PK

Learning Deep CNN Denoiser Prior for Image Restoration (CVPR, 2017) (Matlab)

Home Page: http://openaccess.thecvf.com/content_cvpr_2017/papers/Zhang_Learning_Deep_CNN_CVPR_2017_paper.pdf

MATLAB 99.89% M 0.11%

color-demosaicking deep-model image-deblurring image-denoising image-inpainting image-restoration single-image-super-resolution

ircnn's Introduction

Kai Zhang

[Homepage] [Google Scholar] [ResearchGate] [知乎]

I am currently a postdoctoral researcher at Computer Vision Lab, ETH Zurich, Switzerland.

⚡ News:

Our new work SCUNet for practical image denoising.

⚡ News

We released the testing code of BSRGAN!

🌱 My Repositories

SCUNet	BSRGAN	KAIR	USRNet	DPIR	IRCNN	SRMD	DPSR	DnCNN	FFDNet

The pytorch training and testing codes of SRMD, DnCNN, FFDNet can be found in KAIR.

ircnn's People

Contributors

Stargazers

Watchers

Forkers

issac8huxley tinyloop hengfa hxkwindwizard arasharchor grseb9s dousong yuckfu caomw runngezhang minminaicode liuguoyou zgsxwsdxg huaxinxiao gaobingaobingaobin eraint wujinlonglovezhangmiao1314 supoman shubhampachori12110095 zhangweizhao33 paojianghu yinizhizhu dating99 melisamesan kusiwu hyzcn hxj525279 taixiangjiang chongbingbao davidwuzc wn9081 dadachongzi cuiwenxue chunyu-lin-bjtu gonwalk zumbalamambo simonsroad bygreencn zhouqp631 idoit christinaliang hiredd zoonono sc4dub sdut10523 gao-ricardo hackerjean suntaopku panky8070 caiweicaiwei snowbhr06 nmber5 yuanhaogong liuxiaozhe xiaowen-ttkx pldare fireae briando2005 wang-lizhi xhwxd gorgeousyouth scapeqin sumihui qiaoxie jeromepan yyw15 yanshangongzi klqulei gjlper whu-dft wangke0809 sijiedu gottstaff jwgu shin1919 iamonlyeric ilovedoudou ieee820 jasmeet21 xiaoming3526 syaaaa linsong-xu zhengxinchenee aixigua kaiyue1216 18305169132 datbkhn xbyready chenhongling dalezz1 bilingz benjamesbabala l1ghtzz ganwang uvlrick aceaaaaa zjh1124 iggcaswy oldsuperman xxdm

ircnn's Issues

problems about the size of the training patch and receptive field

Hello, in paper,it said if the size of the training patch is smaller than the receptive
field, the performance would decrease. why? thanks

about your training codes?

Hi @cszn , I have already seen https://github.com/cszn/DnCNN/tree/master/TrainingCodes, but I fail it. Can you give training codes for this project? I find that DnCNN is not IRCNN.

can find drunet_gray.pth in google driver

Hi, thanks for sharing this awesome work, but I can' t find the drunet_gray.pth in the google driver. Did it go by any other names?

about the training code

Hi, sorry for interrupting. Could you please kindly provide the training code as well? Thank you~

For a real-world blurred image

How could I obtain the noise level(Isigma) of a real-world blurred image where the PSF is also unknown?
Is there any algorithm recommended to do that?

Unable to open .mat files

Hello,
I'm trying to run your code on my R2018b matlab on Windows 10.
When I try to load one of your .mat files (e.g. 'Levin09.mat') I get the following error:

Error using load
Unable to read MAT-file C:\Users...\kernels\Levin09.mat. Not a binary MAT-file. Try load -ASCII to read as text.

Nothing I do seems to help.
Any idea how to solve this issue?

pytorch or tensor flow version

hello,
Is there a pytorch or tensor flow version of the same available? If so could you please lead me to it.

To get a blind kernel from a picture

How can i get a blind kernel from a non-specific picture?The non-specific picture is not from the dataset of image restoration.

About the groundtruth images of Deblur_set1

MatConvNet version

Could you provide information as to which MatConvNet version is used by your code? It seems version 1.0 beta 24 crashes.

I run the code for denoise, and the PSNRs and SSIMs are different from those in the paper. For example, the test dataset is DSB68(gray) with the noise level 15.
The mean PSNR I tested was 28.67 and the mean PSNR presented in the paper is 31.63.
Are there any problems with the test code? Or what mistakes I made? If I have made some mistakes, could you help me to solve it? Thanks a lot.

Are there some problems about the trained model?

The dimension of output and input are not consistent. Is there any wrong in the trained model? Or I made some mistake when running the code. But I have not changed anything in the code.

Demo_deblur_real_application

这个demo跑不通，您知道为什么吗？
Error using edgetaper
Expected input number 2, PSF, to be nonzero.

Error in edgetaper>parse_inputs (line 128)
validateattributes(PSF,{'uint8','uint16','int16','double','single'},...

Error in edgetaper (line 34)
[I, PSF, sizeI, classI, sizePSF, numNSdim] = parse_inputs(varargin{:});

Error in Demo_deblur_real_application (line 76)
y = edgetaper(y, k);

About “Demo_inpaint” application

In that application, you use Shepard initialize to fit the picture. But in Shepard function, you use mask as one of its parameter.

I'm very confused about how can we get a picture's "mask.png".Totally we just get a picture without its mask. If your paper just ignores mask's extract just because it is so easy for the paper?

Maybe my problem was so noob.... Sorry about that because I'm new to the CNN...

If you have free time to answer, thank you for your patience!!!

Ｈow do we get the blur kernel?

Dear Author:
Could you please provide some great blind deblur methods to estimating the blur kernel?
Thank you very much!

setting

Sorry, what seetings do I need to configure when I download this files?

Could you help me fix the problem

Error using edgetaper
Expected input number 2, PSF, to be nonzero.

Error in edgetaper>parse_inputs (line 128)
validateattributes(PSF,{'uint8','uint16','int16','double','single'},...

Error in edgetaper (line 34)
[I, PSF, sizeI, classI, sizePSF, numNSdim] = parse_inputs(varargin{:});

Error in Demo_deblur_real_application (line 76)
y = edgetaper(y,k);

For non-matlab models?

Is there a guide to replicate the paper's code in another programming language?

Noisy LR images are generated by Bicubic downsampler instead of Direct downsampler.

How can I get the results I mentioned above.
I notice that your SRMD (CVPR2018) report the IRCNN results using such settings.

Running without GPU

I wanted to use this for my research. I tried to run this in supercomputing clusters which does not have GPUs. It gave errors.

There is a problem with the CUDA driver or with this GPU device. Be sure that you have a supported GPU and that
the latest driver is installed.

How can I run it without GPU?

CUDA Ready GPU is a requirement?(Undefined function or variable 'vl_simplenn_tidy'.)

Dear cszn:
Hi, you are doing really a good job. But I running into and error.
Does it mean I still need a Cuda Support GPU even if I set "use GPU =0" ?
Undefined function or variable 'vl_simplenn_tidy'.

about the the parameter 'isigma'in the 'demo_deblur_real_application'

can the 'Isigma '= 0/255 in the 'Demo_deblur_real_application'?

Demo_deblur_real_application

错误使用 edgetaper
需要的第 2 个输入, PSF, 应为非零。

出错 edgetaper>parse_inputs (line 128)
validateattributes(PSF,{'uint8','uint16','int16','double','single'},...

出错 edgetaper (line 34)
[I, PSF, sizeI, classI, sizePSF, numNSdim] = parse_inputs(varargin{:});

出错 Demo_deblur_real_application (line 76)
y = edgetaper(y, k);

看来您回答别人的要有模糊核，您的样例的模糊核不是文件里就有吗，而且也都出来了，为啥还是有这个问题啊？
希望百忙之中抽空解决一下，谢谢了

About training

What's the training data seem like? Are the input generated by just adding noise to the clean images?
Thanks for your reply ：)

训练数据是否仅仅通过在干净的图片上添加高斯噪声生成？要通过模糊核模糊吗？

kernelsize = ceil(sigma*3)*2+2;
kernel     = fspecial('gaussian',kernelsize,sigma);
blur_HR    = imfilter(im, kernel,'replicate');
LR         = blur_HR(scale/2:scale:end-scale/2,scale/2:scale:end-scale/2,:);

About installation

Hi,
Which OS is required to install and use your works?
Ubuntu or windows? And their version?
Could you please share more information about how to install and use?

Thank you