cszn/dncnn

Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising (TIP, 2017)

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.

dncnn's Issues

About the "bnorm" layers

Hi, I found that the models provided in the folder "model" don't contain "bnorm" layers, while the models that I trained using the provided training codes contain the "bnorm" layers. I want to know if I should remove the "bnorm" layers in my trained models?

Colored images after JPEG deblocking

Is it possible to obtain colored images instead of black and white ones after applying the deblocking procedure?

error 0 in training

Hi, i'm trying to train a model with another set of images and with CPU.
I replaced the images in Train400 and changed in the code GPU = [];
i got error value of 0 and NAN in memory

(parameter memory|2MB (5.6e+05 parameters)|
data memory|NaN (for batch size 128)|
train: epoch 01 dataset 00: 1/5106:error: 0.000000
train: epoch 01 dataset 00: 2/5106:error: 0.000000 )

i cant add image , this is copy paste :

 layer|      0|      1|      2|      3|      4|      5|      6|      7|      8|      9|     10|     11|     12|     13|     14|     15|     16|     17|
  type|  input|   conv|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|
  name|    n/a| layer1| layer2| layer3| layer4| layer5| layer6| layer7| layer8| layer9|layer10|layer11|layer12|layer13|layer14|layer15|layer16|layer17|

----------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|
support| n/a| 3| 1| 3| 1| 1| 3| 1| 1| 3| 1| 1| 3| 1| 1| 3| 1| 1|
filt dim| n/a| 1| n/a| 64| n/a| n/a| 64| n/a| n/a| 64| n/a| n/a| 64| n/a| n/a| 64| n/a| n/a|
filt dilat| n/a| 1| n/a| 1| n/a| n/a| 1| n/a| n/a| 1| n/a| n/a| 1| n/a| n/a| 1| n/a| n/a|
num filts| n/a| 64| n/a| 64| n/a| n/a| 64| n/a| n/a| 64| n/a| n/a| 64| n/a| n/a| 64| n/a| n/a|
stride| n/a| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1|

pad	n/a	1	0	1	0	0	1	0	0	1	0	0	1	0	0	1	0	0
rf size	n/a	3	3	5	5	5	7	7	7	9	9	9	11	11	11	13	13	13
rf offset	n/a	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1
rf stride	n/a	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1
----------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------
data size	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN	NaNxNaN
data depth	NaN	64	64	64	64	64	64	64	64	64	64	64	64	64	64	64	64	64
data num	128	128	128	128	128	128	128	128	128	128	128	128	128	128	128	128	128	128
----------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------	-------
data mem	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN
param mem	n/a	3KB	0B	144KB	1KB	0B	144KB	1KB	0B	144KB	1KB	0B	144KB	1KB	0B	144KB	1KB	0B

Thanks!

Two questions on image denoising

Hi,

I have two questions and hope you can give me some suggestions.

You mentioned "we follow [16] to use 400 images of size 180180 for training" in your paper. I know the 400 images come from the train and test sets of BSD dataset. The original images in BSD have larger size than 180180. I'm wondering how do you crop the 180*180 region in the original image? If it is possible, can you send me a copy? (I have also sent you an email ==!)

Another question is we usually extract patches as the input for the network from the images. So, how do you add the additive gaussian noise? Add on the image and then extract patches from the noisy image, or firstly extract patches from the clean image and then add the noise?

Thanks in advance!

Question about learning rate for bias.

Hi Zhang,

Could you explain the reason not to use bias for the intermediate convolution?
I could see that the first and last convolution define learnRate as lr11, while the other convolutions use lr10.

Thank you.

loss function in DnCNN-Keras

想请问为什么Keras实现里的损失函数里不是以残差R(x)-(x-y)为目标的呢

hello， do you have test model for RGB image？

Question about learning rate decay

Hello! I want to ask a question about the learning rate.
In your paper “Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising”, you mentioned that “The learning rate was decayed exponentially from 1e-1 to 1e-4 for the 50 epochs”.
I wonder how to do this in the code.
Or any others can help me? Thank you very much!!!

BN layers in given model

hi !
Can you please release module with BN layers? (for denoising)
Do you think it should work fine for Poisson noise too?
Thanks !

Error in Demo_test_DnCNN_C (line 57)

Solution please

Error using vl_nnconv
The FILTERS depth does not divide the DATA depth.

Error in vl_simplenn (line 97)
res(i+1).x = vl_nnconv(res(i).x, l.weights{1}, l.weights{2}, ...

Error in Demo_test_DnCNN_C (line 57)
res = vl_simplenn(net,input,[],[],'conserveMemory',true,'mode','test');

vl_simplenn_move.m not found.

Hi, in your demo codes, you included

if useGPU
net = vl_simplenn_move(net, 'gpu') ;
end

However, its not included in your utilities.

how to fine-tuned your model with new data sets

Hi,

How to fined tuned your model with new datasets ??

Why you comment most of your ‘net = vl_simplenn_tidy(net);’

Unbelievable great job！
But why you comment most of your ‘net = vl_simplenn_tidy(net);’ in your code？

Blind Color Denoiser Model

Hi,

I see that you include the pre-trained model for blind denoising on color images under
model/specifics/GD_Color_Blind.mat

I have noticed that this model does not include the parameters of the batch normalization layers. Did you change the weights of the convolution layers after the training by scaling and shifting them with the batch norm parameters or did you just remove the batch norm layers from the model?

I would like to use your model to initialize my network and either I need the full model or I need to have the same training dataset to be able to train it. May I ask you please for your help?

Thanks,
Caner

Are you going to port it to python?

Can I ask a question about sigma setting of blind Gaussian denoising

Hi,
I see that you set the range of the noise levels as σ ∈ [0,55].When I reproduced the blind Gaussian denoising training model, I used random functions to produce sigma. However the effect was not good.
May I ask you please for your help?

Thanks,
Wilson

不加BN层，在BSD68测试集上效果反而更加好

'''
inpt = Input(shape=(40,40,1))
x1 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(inpt)
x2 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x1)
x3 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x2)
x4 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x3)
x5 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x4)
x6 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x5)
x7 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x6)
x8 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x7)
x9 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x8)
x10 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x9)
x11 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x10)
x12 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x11)
x13 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x12)
x14 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x13)
x15 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x14)
x16 = Conv2D(64,(3,3),activation='relu',padding='same',strides=(1,1))(x15)
x17 = Conv2D(1,(3,3),padding='same',strides=(1,1))(x16)
'''
参数和原文一致，只是训练采用了BSD40,当sigma=25,15的时候在测试集BSD68上的psnr比原文要好，而且视觉效果也更棒。

the data in the training data set can be more than 255?

Hi cszn,
I'm trying to train a model with other types of data. the data in the training data set can be more than 255? For example, 0-16384. Error shows 7.6 when I train the model.

Thanks!

Spark

Hello, I wonder if you have tested your code using multi-gpu with matconvnet?

Thank you!

BSD68 RGB dataset

Hi,
I've trained CDnCNN-B in tensorflow and I would like to compare the results but I can't find the RGB version of the BSD68 dataset, can you provide it?

Kind regards,
Michele

MATLAB JPEG encoder?

To generate the input of JPEG Deblocking, use MATLAB JPEG encoder to compress images.
Could you illustrate the MATLAB JPEG encoder? Is JPEG2000 compress algorithm? or other tools?
Thanks

what's exactly the convlution implement?can you offer a reference?

Training with 980Ti data size NAN

Hi there,

I could not train the model with my 980Ti card due to this NAN issue. Does it mean I need a better card or any way to resolve this issues?

Thankyou

Get back the clean image

Hi, is that possible to change the Demo test model in TrainingCodes to get back the clean image instead of the residual image? Or need to re-initialize the model and train the model again?
Thanks!

Reference to non-existent field 'dilate', matconvnet 1.0-beta25

Hi, I got this error when running Demo_test_DnCNN.m with Matlab 2017a, matconvnet 1.0-beta25, Window 10. The details are

Reference to non-existent field 'dilate'.

Error in vl_simplenn (line 303)
'dilate', l.dilate, ...

Error in Demo_test_DnCNN (line 64)
res = vl_simplenn(net,input,[],[],'conserveMemory',true,'mode','test');

Any idea to fix this?
Thank you.

data_size is NAN

I met some difficuties after use vl_compilenn().

>> Demo_Train_model_64_25_Res_Bnorm_Adam
     layer|      0|      1|      2|      3|      4|      5|      6|      7|      8|      9|     10|     11|     12|     13|     14|     15|     16|     17|
      type|  input|   conv|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|
      name|    n/a| layer1| layer2| layer3| layer4| layer5| layer6| layer7| layer8| layer9|layer10|layer11|layer12|layer13|layer14|layer15|layer16|layer17|
----------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|
   support|    n/a|      3|      1|      3|      1|      1|      3|      1|      1|      3|      1|      1|      3|      1|      1|      3|      1|      1|
  filt dim|    n/a|      1|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|
filt dilat|    n/a|      1|    n/a|      1|    n/a|    n/a|      1|    n/a|    n/a|      1|    n/a|    n/a|      1|    n/a|    n/a|      1|    n/a|    n/a|
 num filts|    n/a|     64|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|
    stride|    n/a|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|
       pad|    n/a|      1|      0|      1|      0|      0|      1|      0|      0|      1|      0|      0|      1|      0|      0|      1|      0|      0|
----------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|
   rf size|    n/a|      3|      3|      5|      5|      5|      7|      7|      7|      9|      9|      9|     11|     11|     11|     13|     13|     13|
 rf offset|    n/a|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|
 rf stride|    n/a|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|
----------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|
 data size|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|
data depth|    NaN|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|
  data num|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|
----------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|
  data mem|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|
 param mem|    n/a|    2KB|     0B|  144KB|    1KB|     0B|  144KB|    1KB|     0B|  144KB|    1KB|     0B|  144KB|    1KB|     0B|  144KB|    1KB|     0B|

     layer|     18|     19|     20|     21|     22|     23|     24|     25|     26|     27|     28|     29|     30|     31|     32|     33|     34|     35|
      type|   conv|  bnorm|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|
      name|layer18|layer19|layer20|layer21|layer22|layer23|layer24|layer25|layer26|layer27|layer28|layer29|layer30|layer31|layer32|layer33|layer34|layer35|
----------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|
   support|      3|      1|      1|      3|      1|      1|      3|      1|      1|      3|      1|      1|      3|      1|      1|      3|      1|      1|
  filt dim|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|
filt dilat|      1|    n/a|    n/a|      1|    n/a|    n/a|      1|    n/a|    n/a|      1|    n/a|    n/a|      1|    n/a|    n/a|      1|    n/a|    n/a|
 num filts|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|
    stride|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|
       pad|      1|      0|      0|      1|      0|      0|      1|      0|      0|      1|      0|      0|      1|      0|      0|      1|      0|      0|
----------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|
   rf size|     15|     15|     15|     17|     17|     17|     19|     19|     19|     21|     21|     21|     23|     23|     23|     25|     25|     25|
 rf offset|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|
 rf stride|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|
----------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|
 data size|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|
data depth|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|
  data num|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|
----------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|
  data mem|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|
 param mem|  144KB|    1KB|     0B|  144KB|    1KB|     0B|  144KB|    1KB|     0B|  144KB|    1KB|     0B|  144KB|    1KB|     0B|  144KB|    1KB|     0B|

     layer|     36|     37|     38|     39|     40|     41|     42|     43|     44|     45|     46|     47|     48|     49|
      type|   conv|  bnorm|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|   conv|  bnorm|   relu|   conv|   loss|
      name|layer36|layer37|layer38|layer39|layer40|layer41|layer42|layer43|layer44|layer45|layer46|layer47|layer48|layer49|
----------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|
   support|      3|      1|      1|      3|      1|      1|      3|      1|      1|      3|      1|      1|      3|      1|
  filt dim|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|
filt dilat|      1|    n/a|    n/a|      1|    n/a|    n/a|      1|    n/a|    n/a|      1|    n/a|    n/a|      1|    n/a|
 num filts|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|     64|    n/a|    n/a|      1|    n/a|
    stride|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|
       pad|      1|      0|      0|      1|      0|      0|      1|      0|      0|      1|      0|      0|      1|      0|
----------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|
   rf size|     27|     27|     27|     29|     29|     29|     31|     31|     31|     33|     33|     33|     35|     35|
 rf offset|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|
 rf stride|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|      1|
----------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|
 data size|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|NaNxNaN|
data depth|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|     64|      1|      1|
  data num|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|    128|      1|
----------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|
  data mem|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|    NaN|
 param mem|  144KB|    1KB|     0B|  144KB|    1KB|     0B|  144KB|    1KB|     0B|  144KB|    1KB|     0B|    2KB|     0B|

Does it require GPU to train this network?

I have Matlab in my computer which does not have GPU. Can I train the network on my computer?

color image denoising

Hi czn,

Is there any evaluation on color image denoising task with different noise level?
Like the mean psnr on certain test set?

can i ask you specipic option about uploaded models?

Hi thank you for your work.

can i ask you about specipic train option of uploded models? (like sigma=10.mat, sigma=15.mat)

I think sigma is noise level. but i don't know about how many epoch you train. (in paper train with 50 epoch. is that right?)

Also training codes v1.1 can generate same model? (using adam or SGD?)

Denoise the same picture with different nets

Hi, Thanks for your codes at the first. I am confused by a phenomenon. I have trained 3 nets with sigma = 5,10,15 respectively and I get a noised picture with sigma = 5. Then I used three trained nets to denoised the same picture, but I find the net trained with sigma = 5 performs worse than other two nets. And I find the nets with sigma = 15 performs best. I am confused by this phenomenon, do you know the reason?

Overfitting problem

Hi, do we need consider over-fitting problem when we fine-tune the parameters to reproduce the results? If is necessary, can I use PSNR as evaluation metric to evaluate the training loss and test loss?

Hello

hello, i want use the tensorflow to make a dncnn, so, i need train the image, but in this code, just like have the target image(without noise; -Train400), i also need the input image(with noise),can you seed to me? thanks
my email address [email protected]

strange output from the DnCNN denosing

Hi,
I tried to run the Demo_test_DnCNN.m, but the output image looks very strange, there are several vertical stripes in the image, and the stripes exist in all output images. and I also tried the Demo_test_FDnCNN_Gray.m, still the same issue.

About the usage of DnCNN with matconvnet-1.0-beta24 on linux

Thanks so much for sharing your code. I tried to use it and find two problems:

1. The slash in the path on Linux is different from Windows:

     folderTest  = 'testsets\Set12'; %%% test dataset

2. For "matconvnet" version later than 20, some modifications need to be made.
   A bug " missing field 'dilate' in layers" shows up. I tried to fix this by using

   net = vl_simplenn_tidy(net)
   net = vl_simplenn_move(net,'gpu');

before

   res    = vl_simplenn(net,input,[],[],'conserveMemory',true,'mode','test');

Now it works.

Is there anyone who has implemented this code? And made his own changes?

I am almost done with the training code...but i want to make some changes. is there any person who can help me. This is my email address. please feel free to contact me....

[email protected]

sorry can i ask your one more question?

in the paper DnCNN-B is blind test about noise level.

that means, you train image given random noise in [0,55]?

or some image give small noise, other give big noise?

can i ask you specific condition or code about DnCNN-B train?

Struggle with PyTorch version code

I'm afraid I might bother you, but it seems that the PyTorch implementation cannot reproduce the expected result.
When I tried to do so, I met following issues......:

When I run train/test code(without touching any setting),

• DnCNN-S for noise level=25 made PSNR=29.24 in <30 epoch
• DnCNN-S for noise level=15 got stuck at PSNR=23.xx. I tried this many times but didn't work.
• DnCNN-S for noise level=50 made 25.9x in <50 epoch, but couldn't reach near 26.23 at all.
  There's multi_step scheduling, however I forced it to use smaller learning rate after the convergence, and it didn't help.
  +) I tried some modifications on the model, and still the problem remained. Just <0.8dB changed. So think it is not about model.
  +) I also tried to change the optimizer and lr scheduling exactly same as it is mentioned on the original paper, but that made it even worse.

Is it right to add Gaussian noise and don't clip it? I thought it must be clipped according to [0., 255.] (uint8) or [0., 1.] (float32) because in real case(of course it is not completely 'real' 'cause it's AWGN) the corrupted image would be in [0., 255.] range.
How much does the clipping lowers its performance, or is it better? I wanted to test it myself but... like I mentioned in 1., the baseline model isn't working correctly, so... If you have already tried it, could you let me know?

Understanding vl_nnloss.m

Hi there,

I am trying to modify the loss function in DnCNN, but I am not quite sure how to do it.
Could you explain a little more about your vl_nnloss?
I understand that

t = ((X-c).^2)/2;
Y = sum(t(:))/size(X,4); % reconstruction error per sample;

calculates the L2 loss function.
How about the second part?

Y = bsxfun(@minus,X,c).*dzdy;

why did you use bsxfun and minus? and could you explain the meaning of dzdy?

Thank you very much,
Best regards,
Atena

Demo_test_DnCNN.m

Hi!I meet a problem when I run 'Demo_test_DnCNN.m'.
System error as:
错误使用 vl_nnconv
An input is not a numeric array (or GPU support not compiled).

出错 vl_simplenn (line 300)
res(i+1).x = vl_nnconv(res(i).x, l.weights{1}, l.weights{2}, ...

出错 Demo_test_DnCNN (line 64)
res = vl_simplenn(net,input,[],[],'conserveMemory',true,'mode','test');
I‘m a beginner. Could you tell me how can I solve it,or anyone knows the answer,please tell me ,thank you very much!

Pytorch model doesn't produce reported PSNR scores.

I tried to run the Main_test.py code. But I am unable to get the reported scores for both Set 68 and Set 12 test datasets.

model with sigma 5 or 1

Hi!

Do you maybe have a net you trained with sigma 1 or 5?

Thanks!
Sivan

receptive field size

Hi, based on my understanding, receptive field size is same as the convolutional filter size, according to model initialization under the training code, filter size is 3x3 for all the layers, so receptive field size is 3x3 for all layers regardless the depth of the network.
However, the report says: "the receptive field of DnCNN with depth of d should be (2d + 1) x (2d + 1). Thus, for Gaussian denoising with a certain noise level, we set the receptive field size of DnCNN to 35x35 with the corresponding depth of 17."
Do you mind further explain it so that I know where I am wrong?
Thanks!

why your models trained already are so small,though they have a large depth? they're very memory-friendly,I'm wondering how you get there

Forward:out of memory

Hi cszn,
Meet you again~
I have 11GB available in GPU, Why I got out of memory ERROR?
I im using a BMP(3968*2976) to feed the NET.

Where can I get train400 dataset?

The train400 data from BSD dataset is used for training it, following those works:

Chen, Yunjin, and Thomas Pock. "Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration." IEEE transactions on pattern analysis and machine intelligence 39.6 (2017): 1256-1272.
Schmidt, Uwe, and Stefan Roth. "Shrinkage fields for effective image restoration." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2014.

and well, I couldn't find a list of these 400 images(and also tried to find the exact image files like BSD68, but couldn't).
What are these and how can I get them?

error in main_train.py

after successfully running the data_generator.py code. when i run the code of main_train.py it shows the following error

File "C:\Users\Sufian\Anaconda3\lib\site-packages\spyder\utils\site\sitecustomize.py", line 102, in execfile
exec(compile(f.read(), filename, 'exec'), namespace)

File "C:/Users/Sufian/Desktop/dncnn_keras/main_train.py", line 69
x = BatchNormalization(axis=3, momentum=0.0,epsilon=0.0001, name = 'bn'+str(layer_count))(x)
^
TabError: inconsistent use of tabs and spaces in indentation

can i ask you about last layer?

Hi I have bin interest about your method.

can i ask you one question?

I don't under stand about last layer.

last layer got 64 filtered image at previous layer and use 64 filters for convolution.

but how reconstruct one image?

reconstruct image pixel is sum of each 64 filter image pixel?

MLP code for sigma = 50

Hi,

I saw in your paper that you have compared against MLP with sigma=50. Please can you direct me to the link or share the code for MLP sigma=50.

Waiting for your kind response.

Regard,
Saeed

CBSD432 dataset for CDnCNN-B

For training CDnCNN-B, did you extract patches from the full-sized CBSD432? Or from 180x180 crops? If the latter is true, can you provide the link to the cropped CBSD432 dataset.

Quesitons about the training process.