Continuous Patch Stitching for Block-wise Image Compression

Zifu Zhang, Shengxi Li, Henan Liu, Mai Xu, Ce Zhu
🥳 This work is accepted by IEEE Signal Processing Letter.

📖 Table Of Contents

• 💻 Train
• ⚡ Inference
• ❤️ Acknowledgement
• 📋 Citation

🔧 Requirements

- conda create -n cps python=3.8
- conda activate cps
- pip install compressai
- pip install pyyaml
- pip install easydict
- pip install tensorboard

💻 Train

1. Prepare the training dataset and change the path in ./configs/xxx.yaml.

   save_folder
   ├── train # train dataset
   └── test # test dataset

2. Run the following command.

   python train.py --config configs/cps_1_lambda00018.yaml

⚡ Inference

1. Download the checkpoints into ./experiment.

   lambda	Link
   0.0018	cps_6
   0.0035	cps_5
   0.00625	cps_4
   0.013	cps_3
   0.025	cps_2
   0.0483	cps_1

2. Download test datasets.

• Kodak
• CLIC2020_val
• DIV8k

3. Run the following command.

   python test.py checkpoint --dataset Kodak -a cps-v4 \
   -p experiment/cps_1_lambda00018_v4_best_loss.pth.tar \
   -d output/cps_lambda00018_v4_patch256_Kodak \
   --config configs/cps_1_lambda00018.yaml --save-image --per-image --patch --cuda

❤️ Acknowledgement

This work is based on Compressai, thanks to the invaluable contributions.

📋 Citation

Please cite us if our work is useful for your research.

@article{zhang2025continuous,
  title={Continuous Patch Stitching for Block-wise Image Compression},
  author={Zhang, Zifu and Li, Shengxi and Liu, Henan and Xu, Mai and Zhu, Ce},
  journal={arXiv preprint arXiv:2502.16795},
  year={2025}
}