Paper for Journal of Signal Processing (2024 January)
Joint Training of Noisy Image Patch and Impulse Response of Low-Pass Filter in CNN for Image Denoising
May Thet Tun, Yosuke Sugiura, Tetsuya Shimamura

% A PyTorch implementation of FFDNet image Denoising.
====================================================
## Dependencies
* [PyTorch v0.3.1](http://pytorch.org/)
* [scikit-image](http://scikit-image.org/)
* [torchvision](https://github.com/pytorch/vision)
* [OpenCV](https://pypi.org/project/opencv-python/)
* [HDF5](http://www.h5py.org/)
* [tensorboardX](https://github.com/lanpa/tensorboard-pytorch)


## Dataset Information
======================
For Color image;
Training Dataset = CBSD500
Testing Dataset = CBSD68, Kodak24,RNI15 (realistic testing dataset)

For Grayscale image;
Training Dataset = train
Testing Dataset = Set12, BSD68,RNI6 (realistic testing dataset)

## Usage
=========
(1) Create prepare patches
=========================
For Color image;
python prepare_patches.py --aug_times 1;
For Grayscale image;
python prepare_patches.py --aug_times 1 --gray;

The default size of the stride is set to 20, while augmentation time can be chosen (with the default being 1).
Users have the flexibility to specify patch size, FFDNet recommends grayscale for 70 and color for 50.

A model can be trained after having built the training and validation databases.
(i.e. *train_rgb.h5* and *val_rgb.h5* for color denoising, and *train_gray.h5*
and *val_gray.h5* for grayscale denoising).

(2) Model Training
==================
For Color image;
python OrgTrain.py --epochs 80 --noiseIntL 0 75 --val_noiseL 25;
For Grayscale image;
python OrgTrain.py --epochs 80 --noiseIntL 0 75 --val_noiseL 25 --gray;

* By default, models are trained for values of noise in [0, 75] and
* noise added at validation is set to 25 according to the recommendation of FFDNet.


(3) Model Testing
=================
For Grayscale image,
python OrgTest.py --noise_sigma 25 --no_of_channel 1 --add_noise true;
For Color image,
python OrgTest.py --noise_sigma 25 --no_of_channel 3 --add_noise true;

* *add_noise* can be set to *false* if the input image is already noisy. For example, RNI6 and RNI15
are realistic noisy image. For realistic dataset RNI6 and RNI15 add_noise= false;
For example,
python OrgTest.py --noise_sigma 25 --no_of_channel 3 --add_noise false;

* By default, models are trained for values of noise in [0, 75] in --noiseIntL parameter and
* noise added at validation is set to 25 in -val_noiseL parameter
** A previous training can be resumed passing the [--resume_training] flag at the end of the model training comment.

** Modify the training and testing of the file names for DLPF, BLPF, GLPF, and SLPF to different requirements.
For power spectrum sparsity based Low Pass Filter, SLPFTrain.py and SLPFTest.py
For Direct ( Ideal ) Low pass filter , DLPFTrain.py and DLPFTest.py
For Gaussian Low pass filter , GLPFTrain.py and GLPFTest.py
For Butterworth Low pass filter , BLPFTrain.py and BLPFTest.py