IDKiro / CBDNet-pytorch

Toward Convolutional Blind Denoising of Real Photograph

CBDNet-pytorch

It's an unofficial PyTorch implementation of CBDNet.

We used higher quality real and synthetic datasets for training and achieved better performance on DND.

CBDNet in MATLAB

CBDNet in Tensorflow

Quick Start

Download the dataset and pretrained model from GoogleDrive.

Extract the files to data folder and save_model folder as follow:

~/
  data/
    SIDD_train/
      ... (scene id)
    Syn_train/
      ... (id)
    DND/
      images_srgb/
        ... (mat files)
      ... (mat files)
  save_model/
    checkpoint.pth.tar

Train the model:

python train.py

Predict using the trained model:

python predict.py input_filename output_filename

Network Structure

Realistic Noise Model

Given a clean image x, the realistic noise model can be represented as:

Where y is the noisy image, f(.) is the CRF function and the irradiance , M(.) represents the function that convert sRGB image to Bayer image and DM(.) represents the demosaicing function.

If considering denosing on compressed images,

Result