kepengxu / fidelity_controllable_compression

Implemention of Fidelity Controllable Extreme Image Compression with GAN

Fidelity-Controllable Extreme Image Compression with Generative Adversarial Networks

This repository is a PyTorch implemention of following paper:
Fidelity-Controllable Extreme Image Compression with Generative Adversarial Networks
arxiv url (https://arxiv.org/abs/2008.10314)
Shoma Iwai, Tomo Miyazaki, Yoshihiro Sugaya, and Shinichiro Omachi

Our Environment

    Python==3.6.9
    pytorch==1.0.0
    scipy==1.3.2
    numpy==1.17.4
    opencv-python==4.1.1.26
    tqdm==4.38.0

Pretrained Model

Download our pretrained model and unzip it.
ckpt_model*_mse.pth is trained in the first stage, and ckpt_model*_gan.pth is fine-tuned in the second stage. These two models share the same encoder.

model name	Average bitrate (Kodak)
ckpt_model1_*.pth	0.0299 bpp
ckpt_model2_*.pth	0.0623 bpp

Test

    python compress.py MODEL_PATH IMAGE_PATH

Compressed files will be stored at outputs/binary.

    python decompress.py MODEL_PATH BINARY_PATH

For example,

    python compress.py checkpoints/ckpt_model1_gan.pth images/
    python decompress.py checkpoints/ckpt_model1_gan.pth outputs/binary

Network Interpolation

If you want to use network interpolation, run decompress_netinterp.py.

    python decompress_netinterp.py MODEL_PATH MODEL_PATH2 ALPHA BINARY_PATH

For example,

    python compress.py checkpoints/ckpt_model2_gan.pth images/
    python decompress_netinterp.py checkpoints/ckpt_model2_gan.pth checkpoints/ckpt_model2_mse.pth 0.8 outputs/binary

You can balance the trade-off between distortion and perception by changing alpha.

Acknowledgments

We thank nayuki and Jorge Pessoa for the code of atirhmetic coding and GDN, respectively.