sorry can i ask your one more question?

cszn / DnCNN

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

https://cszn.github.io/

1.47k stars 539 forks source link

sorry can i ask your one more question? #25

Closed sooyeonshin closed 6 years ago

sooyeonshin commented 7 years ago

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?

cszn commented 7 years ago

It is a random value from [0, 55] for each patch. DnCNN-B is mainly used to show the capacity of CNN. It does not generalize well to real noisy images with non-AWGN.

Possible solution to enhance the generalization to real noise includes:

using discrete noise levels such as {5,10,15,..,50,55} in training;
adopting early stopping during training;
designing robust architecture.

Here is my new non-blind denoising work FFDNet.

cszn commented 6 years ago

Try the following code to synthesize the Gaussian noise with noise level in the range of [0, 55].

sigma   = 55;
sigmas  = (sigma*rand(1,size(inputs,4),'single'))/255;
labels  = randn(size(inputs),'single').*repmat(permute(sigmas,[3 4 1 2]),[size(inputs,1),size(inputs,2)]);

Here is my new singe image super-resolution (SISR) work SRMD.