DnCNN-PyTorch

This is a PyTorch implementation of the TIP2017 paper Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising. The author's MATLAB implementation is here.

This code was written with PyTorch<0.4, but most people must be using PyTorch>=0.4 today. Migrating the code is easy. Please refer to PyTorch 0.4.0 Migration Guide.

How to run

1. Dependences

• PyTorch(<0.4)
• torchvision
• OpenCV for Python
• HDF5 for Python
• tensorboardX (TensorBoard for PyTorch)

2. Train DnCNN-S (DnCNN with known noise level)

python train.py \
  --preprocess True \
  --num_of_layers 17 \
  --mode S \
  --noiseL 25 \
  --val_noiseL 25

NOTE

• If you've already built the training and validation dataset (i.e. train.h5 & val.h5 files), set preprocess to be False.
• According to the paper, DnCNN-S has 17 layers.
• noiseL is used for training and val_noiseL is used for validation. They should be set to the same value for unbiased validation. You can set whatever noise level you need.

3. Train DnCNN-B (DnCNN with blind noise level)

python train.py \
  --preprocess True \
  --num_of_layers 20 \
  --mode B \
  --val_noiseL 25

NOTE

• If you've already built the training and validation dataset (i.e. train.h5 & val.h5 files), set preprocess to be False.
• According to the paper, DnCNN-B has 20 layers.
• noiseL is ingnored when training DnCNN-B. You can set val_noiseL to whatever you need.

4. Test

python test.py \
  --num_of_layers 17 \
  --logdir logs/DnCNN-S-15 \
  --test_data Set12 \
  --test_noiseL 15

NOTE

• Set num_of_layers to be 17 when testing DnCNN-S models. Set num_of_layers to be 20 when testing DnCNN-B model.
• test_data can be Set12 or Set68.
• test_noiseL is used for testing. This should be set according to which model your want to test (i.e. logdir).

Test Results

BSD68 Average RSNR

Set12 Average PSNR

Tricks useful for boosting performance

• Parameter initialization:
  Use kaiming_normal initialization for Conv; Pay attention to the initialization of BatchNorm

def weights_init_kaiming(m):
    classname = m.__class__.__name__
    if classname.find('Conv') != -1:
        nn.init.kaiming_normal(m.weight.data, a=0, mode='fan_in')
    elif classname.find('Linear') != -1:
        nn.init.kaiming_normal(m.weight.data, a=0, mode='fan_in')
    elif classname.find('BatchNorm') != -1:
        m.weight.data.normal_(mean=0, std=math.sqrt(2./9./64.)).clamp_(-0.025,0.025)
        nn.init.constant(m.bias.data, 0.0)

• The definition of loss function
  Set size_average to be False when defining the loss function. When size_average=True, the pixel-wise average will be computed, but what we need is sample-wise average.

criterion = nn.MSELoss(size_average=False)

The computation of loss will be like:

loss = criterion(out_train, noise) / (imgn_train.size()[0]*2)

where we divide the sum over one batch of samples by 2N, with N being # samples.