xhwxd / dbsn Goto Github PK

Hi, may I ask we need to pre-train only for the RGB image or both grey and RGB images all need pre-train μ first by using L2 Loss, then pre-train σ_μ and σ_n by using the Pre-trained loss ?

Hi, when I used the provided pre-trained model with default arguments and code, I found that some parameters are not correct which leads to no corresponding parameters, e.g. rgb_gaussian_nL25.pth..

Thank you for any helpful suggestions!

When I train the gray model, the loss becomes negative. It is very confusing.

Hi, in README you mentioned that

For D-BSN，we suggest to pre-train μ first by using L2 Loss, then pre-train σ_μ and σ_n by using the Pre-trained loss (./net/losses.py), finally fine-tune the whole framework.

In your .sh script I noticed the dataset for pre-training μ is actually synthetic data (by adding poisson-gaussian noise to the clean images). My question is, will D-BSN be degraded or improved if use real noisy dataset to pre-train μ?

Thanks.

I am very interested in the actual effect of the paper. I tried to run rgb_test.sh, and the final result was a picture of synthesized noise. Can you provide a demo of denoising? gratitude

Hi,

I met a problem when using "rgb_gaussian_nL50.pth" model. I entered "dbsn_color" folder and modified rgb_test.sh to

But there is an error occured like that:

So I traversed the key in "rgb_gaussian_nL50.pth" model, it shows:

It seems that a dtcn model is in this checkpoint. Could you help me run rgb_gaussian model? Thank you very much!

I wonder if we could have access to knowledge distillation code for rgb image? Thank you very much.

Hi, it's glad to see such a creative work on unsupervised denoising area. And could you please offer the trained model for D-BSN?