DOGNet

JPEG Artifact Reduction Based on Deformable Offset Gating Network Controlled by a Variational Autoencoder

Prerequisites

• python 3.7
• pytorch 1.13

Dataset Preparation

Download Training Dataset DIV2K and Flickr2K.

You should have following directory structure:

dataset
    |-- DIV2K_train
    |-- Flickr2K
        |-- 000001.png
        |-- ...
    |-- ...

Training

• Run train.py

--gpu : GPU Index. If you want to use mutliple GPUs, use --mgpu True
--exp_name : Name of the experiment
--train_dataset : Name of the dataset

python train.py --gpu [GPU INDEX] --exp_name [EXP_NAME] --train_dataset [Datatset Name]

For multiple GPUs

python train.py --gpu [GPU INDEXs] --mgpu True --exp_name [EXP_NAME] --train_dataset [Datatset Name]

Experimental Results

Results on Grayscale JPEG Artifacts Removal

Results on Color JPEG Artifacts Removal

Additional Visualized Results on Real-World Images

Restoration of JPEG images in a real-world scenario; Images are taken by iPhone12 and then uploaded to an SNS

Citation

@article{yoon2023dognet,
  title={JPEG Artifact Reduction Based on Deformable Offset Gating Network Controlled by a Variational Autoencoder},
  author={Yoon, JeongHwan and Cho, Nam Ik},
  journal={IEEE Open Access},
  year={2023},
  publisher={IEEE}
}