How to know the noise intensity of an image, if test a good image and what result will get

Hi,

Thanks for your amazing work.

I would like to confirm the exact training set of FFDNet, in your paper, you stated that "We collected a large dataset of source images, including 400 BSD images, 400 images selected from the validation set of ImageNet [53], and the 4,744 images from the Waterloo Exploration Database [54]".

I would like to ask where could I find the 400 BSD images and the 400 images selected from the validation set of ImageNet?

Hi, I was wondering if anyone can help with implementing non-uniform noise level map in the Pytorch version. It is only implemented for the MATLAB version in the existing codes. Thanks!

Hi there,

I just run the demo_awgn_gray and got this error

Demo_AWGN_Gray
Error using vl_simplenn (line 377)
Unknown layer type 'SubP'.
Error in Demo_AWGN_Gray (line 88)
res = vl_simplenn(net,input,[],[],'conserveMemory',true,'mode','test');

Any idea to fix it?

Best regards,

Hi,

I tried to use an abstract interpretation based verifier to over-approximate and analyze your network FFDNet.
But the network contains 175,616 neurons, 15 layers to denoise even mnist image.
The verifier can't scale up to handle this large and deep network.
Even though I don't think it is possible, but still want to check if you have a smaller size network?
Thanks!

As for me , this article has nearly the same idea with Demosaic area , especially with paper : Deep Joint Demosaic and Denoise,which is from Fredo

the fast speed just comes from packing the original image to a smaller resolution

the good noise adaptive ability comes from the addtional noise estimation input

Hello, it is more of a question than an issue.
I was reading recent papers on image denoising and have read DnCNN and FFDNet papers. I was curious as to why you have not compared FFDNet with RBDN -- it was mentioned in related work in FFDNet. Is the comparison going to be there in final version of FFDNet -- as RBDN is also quite recent.

Thanks,
Touqeer

@cszn hi guys,thanks for your great job,it performs better than CBDnet ,dncnn in my data,i want transplants this algorithm to embedded platform. i have trouble in convert this training result .pth to .onnx,can you give me some advice,looking for you replay,thanks a lot.

The method could be framework for other image processing like Demosaic, HDRNet and Learning to see in the dark...

一个统一的处理架构针对几乎连续的场景，这是Deep Learning可以真正工业化的方法。
赞！

This is not an issue. Appciate your innovation!

In your file model_train.m, you multiply the sigmas with an array of random numbers. Since randn produces negative numbers such as -3.14, the sigma will exceed 75 in the image. My question is that why do you multiply the sigmas with that array

K = randi(8);
labels = imdb.HRlabels(:,:,:,batch);
labels = data_augmentation(labels,K);
sigma_max = 75;
sigmas = (rand(1,size(labels,4))*sigma_max)/255;
<-------------- This line! ------------------>
new_arr = bsxfun(@times,randn(size(labels)), reshape(sigmas,[1,1,1,size(labels,4)]));
inputs = labels + new_arr;

@cszn Can you release the code for training ? It is a long time....

你好，我最近在看你写的FFDNet的论文，想自己训练一个模型，我看到代码里面黑白图像和彩色图像是分开来处理的，请问是这样子的吗？还是我看错了？谢谢

I am having trouble while using your training code. The error is not decreasing after a certain threshold (2.8). Can you tell me how did you manage to train your model.

Thanks for publishing the code. The FFDNet shows very good performance on real noise images as denoted as FFDNet+. But I cannot find any description about it. Could you add some explanation on it?

All the best.

I use your matlab version and pytorch for test ffdnet color model, matlab version's time is about 20ms, but pytorch version's time is about 1s. Cuda version is 8.0 and cudnn is 6.0.21. Can you have some ideas about this? What is your testing time and pytorch version?

Hi,

For AWGN images, you mention in the paper that the noise level map should take uniform element \sigma.
May I know why the noise level map is set in this way for AWGN?
And is it ok to denoise AWGN image with non-uniform noise level map?

Thanks,
Yuyi

Hi @ @cszn
This project and paper are great.
I want to test spatially variant noise images with noise level mask (Fig5 in arxiv paper), can you share me the test codes?
Thanks a lot

Hi @ @cszn
This project is such good. But, I want to train mu own dataset, can you share me the training codes? Thanks a lot

can you share training dataset about color images,thanks

Each training data with an uniform map? (or is there some training data with spatially variant noise map in training stage?)

thank you 😄

Thanks for your work. But I have a question about how can num_input_channels=15 in training step and 3 in test step?

I use the Train400 dataset for FFDNet-master\TrainingCodes\FFDNet_TrainingCodes_v1.0\Demo_Train_FFDNet_gray.m.
The training looks no problem, outputs like below -
train: epoch 499 : 1/ 4: loss: 2.3612
train: epoch 499 : 2/ 4: loss: 3.2255
train: epoch 499 : 3/ 4: loss: 3.5821
train: epoch 499 : 4/ 4: loss: 3.0635
train: epoch 500 : 1/ 4: loss: 3.9296
train: epoch 500 : 2/ 4: loss: 4.1511
train: epoch 500 : 3/ 4: loss: 2.4029
train: epoch 500 : 4/ 4: loss: 2.2997
But I did not see the final trained model. Could you please help to give some recommendation that what should I do?

First of all, thank you for sharing your scientific progress with the github community.

I opened this issue to ask and keep a track in a possible keras implementation of this work.
Is there any plan to release a keras version in a near future?

`
imgn_train = img_train + noise

Create input Variables

img_train = Variable(img_train.cuda())
imgn_train = Variable(imgn_train.cuda())
noise = Variable(noise.cuda())
stdn_var = Variable(torch.cuda.FloatTensor(stdn))

Evaluate model and optimize it

out_train = model(imgn_train, stdn_var)
loss = criterion(out_train, noise) / (imgn_train.size()[0] * 2)
loss.backward()
optimizer.step()
`

这个代码是你在首页推荐的，http://www.ipol.im/pub/pre/231/ 里面的代码。

你看他计算 loss 的时候，是将模型的输出与 noise 进行计算，也就是说模型预测的是 噪声，怎么感觉跟你的论文说的不一样啊

please! When I use the pytorch model, I meet an error.
I want to denoise an image using a one of the pretrained models

python test_ffdnet_ipol.py
--input input.png
--noise_sigma 25
--add_noise True

I have:

Testing FFDNet model

Parameters:
add_noise: True
input: 111.png
suffix:
noise_sigma: 0.09803921568627451
dont_save_results: False
no_gpu: False
cuda: True

rgb: False
im shape: (518, 774)
Loading model ...

Process finished with exit code 0
I have no use to get anything, but I have configured the libraries that need to be configured.
Then I debugged it and when I debugged this statement, the project exited.
if args['cuda']:
state_dict = torch.load(model_fn)
device_ids = [0]
model = nn.DataParallel(net, device_ids=device_ids).cuda()

I want to know why I can't use the training model?@cszn

I noticed that you mentioned RGB version of AWGN spatially variant denoising, but I don't find it in this repo. Is there RGB version AWGN spatially variant denoising pre-trained model and test demo? Many thanks!

I can't run the demo, because 'vl_simplenn_tidy' don‘t exist. And it seems that some functions don't exist too.

Hi @ @cszn
I was very impressed from your works including good peformance of DND set.
I heared that your sigmas are manually selected,
could you share me the applied sigmas for each images at the DND?

Thanks a lot

Hello Sir.

I tried to get "main_train_ffdnet.py" .
When I clicked link, but I couldn't get this file.

How to get this file..

Thanks,,,
Edward Cho.

Is there is any code that fine-tuned your model with new datasets and compare the results ??

Thanks

Hi, initially I would like to thank you for your very good work.
Then I have a question: I've images with more than one channel and the noise level is different in different channels.
How should be setted the noise level map?
Thanks for your help.