Deep Residual Autoencoder for Real Image Denoising

Details about the project and demo images can be found at project website.

Dependencies

• torch>=1.6.0
• torchvision>=0.7.0
• Pillow~=7.2.0
• numpy~=1.19.2
• piq~=0.5.1
• pytorch_msssim~=0.2.1
• torchsummary~=1.5.1
• tensorboard~=2.3.0
• prefetch_generator~=1.0.1
• tqdm~=4.50.2
• matplotlib~=3.3.2

Installation

1- Clone this repo.

git clone https://github.com/yilmazdoga/Deep_Residual_Autoencoder_for_Real_Image_Denoising.git

2- Go to the project directory.

cd  <CLONED_PROJECT_DIR>

3- Create a virtual environment.

python3 -m venv venv

4- Activate the virtual environment.

source venv/bin/activate

5- Install required packages.

pip install -r requirements.txt

At this point you should be able to use the pretrained models to denoise a given image. However, if you want to train the model on your machine or run the test script on the validation data continue the installation with the following steps.

6- Download the Smartphone Image Denoising Dataset (SIDD) medium sRGB version from here.

7- Unzip the downloaded file.

8- Move the Data folder inside the unzipped file to PROJECT_ROOT/datasets/SIDD_Dataset/SIDD_Medium_sRGB/

9- Download the SIDD Noisy and Ground-Truth sRGB validation data from here.

10- Move the downloaded files to PROJECT_ROOT/datasets/SIDD_Dataset/SIDD_Validation/

At this point you should be able to use all of the functionalities.

Usage

Training

To start a training run the following command with desired options.

python3 train.py YOUR_OPTIONS

Example:

python3 train.py --model AEv2_0 --loss-func L1_MSSSIM --augs MixUp_Tensor --epochs 200 --batch-size 24 --patch-size 128 --gpus 2

For further options inspect options/base_options and options/base_options.

Testing

To denoise an image run the following command with desired options.

python3 test.py --noisy-image-path PATH_TO_YOUR_NOISY_IMAGE --gt-image-path PATH_TO_YOUR_GT_IMAGE --YOUR_MODEL_OPTIONS

If you dont have Ground-Truth (GT) image you can give noisy image as GT image but doing so will result a wrong SSIM and PSNR calculation.

To run the test code over the validation dataset run the following command. To use this feature make sure that no Noisy or GT input image path is provided.

python3 test.py --YOUR_MODEL_OPTIONS

Contribute

Please feel free to open an issue or to send an e-mail to doga.yilmaz.11481@ozu.edu.tr