Awesome

Scalable Near-lossless Image Compression

Source code of our CVPR 2021 paper "Learning Scalable ℓ_∞-constrained Near-lossless Image Compression via Joint Lossy Image and Residual Compression".

Usage

To run the code requires Python 3.6 and TensorFlow 1.15.

pip install tensorflow-compression==1.3
pip install range-coder

The main.py provides compress and decompress functions, and an example to encode/decode ./test_patch/kodim05_p128.png. Users can test their own images.

The model ckp_003 can be downloaded from Baidu Netdisk, with access code snic.

Please note: the current implementation is not optimized for speed. Residual compression is slow. We are working on the fast version :computer:.

Citation

@InProceedings{Bai_2021_SNIC,
  title={Learning Scalable $\ell_\infty$-constrained Near-lossless Image Compression via Joint Lossy Image and Residual Compression},
  author={Bai, Yuanchao and Liu, Xianming and Zuo, Wangmeng and Wang, Yaowei and Ji, Xiangyang},
  booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2021}
}