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

Blind Color Denoiser Model #7

Closed hazirbas closed 7 years ago

hazirbas commented 7 years ago

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

cszn commented 7 years ago

1) About BN. See Merge_Bnorm_Demo.m. 2) About initialization for color image. Change the setting of the first and last Conv layer in DnCNN_init_model_64_25_Res_Bnorm_Adam.m. 3) About other color models. I also added the specific color models, see specifics_color. 4) About training datasets. See datasets.

hazirbas commented 7 years ago

Thanks! This will do the magic ;)

xiaofeiwu commented 7 years ago

It seems that the codes in DnCNN_TrainingCodes_v1.1 are still for grayscale images, while I can't find the codes for color image denoising. Would you give more details on how to train a model for color image denoising?