unet_artifact_removal

A CNN-based model that could remove jpeg artifacts from a given jpeg image.

• For details on artifact removal, purpose, model architecture and training, please refer this article.

Training and Inference:

• Tensorflow version 2.1 is required to execute training and inference of the model.

• Installation of necessay libraries can be done using the command given below:

pip install -r requirements.txt

• Download the dataset for training and testing by executing the command given below:

gdown 1Bppd8WW2N7Ji2BVizDINBkbuaZg9lkUY
unzip jpeg_dataset.zip
rm jpeg_dataset.zip

• Please make sure that the dataset is present in the same directory as the repository.

• For training, execute command given below:

mkdir outputs
python train.py --steps 20000

• This will train the model for 20,000 steps and will output a folder named "UNET_month_day_hour_minute_second".

• Folder named outputs would be created while training that has model's output results at every step.

• Another folder named __outputs will be given out that has checkpoint and summary details, please do not delete this folder.

• To check model's performance on some image from the dataset, please execute the command given below:

python infer.py --ckpt_path path/to/UNET_month_day_hour_minute_second/ckpt-20000

• This will output two image files, one would be the downgraded image and another would be the artifact removed image.

Code and concept Credits: ppooiiuuyh, ShakedDovrat