Fidelity-Controllable Extreme Image Compression with Generative Adversarial Networks

This repository is a PyTorch implementation of following paper:
Fidelity-Controllable Extreme Image Compression with Generative Adversarial Networks
ICPR2020 Accepted (arxiv)
Shoma Iwai, Tomo Miyazaki, Yoshihiro Sugaya, and Shinichiro Omachi

Our Environment (updated 2022/09/10)

    Python==3.6.9
    pytorch==1.0.0
    scipy==1.3.2
    numpy==1.17.4
    opencv-python==4.1.1.26
    range-coder==1.1
    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	0.0300 bpp
ckpt_model2_*.pth	0.0624 bpp

Train

    python train.py ./config/default.json

Learned weights will be stored at checkpoint/{exp}/model. Note that, train.py is a very simple training code with minimal functionality.

Test (updated 2022/09/10)

Encoding

    python codec.py encode -c CONFIG_PATH -m MODEL_PATH -i IMAGE_PATH -o BIN_PATH

Decoding

    python codec.py decode -c CONFIG_PATH -m MODEL_PATH -i BIN_PATH -o SAVE_IMG_PATH

For example,

    python codec.py encode -c config/default.json -m checkpoint/ckpt_model1_gan.pth -i images/kodim01.png -o outputs/binary/kodim01.pth
    python codec.py decode -c config/default.json -m checkpoint/ckpt_model1_gan.pth -i outputs/binary/kodim01.pth -o outputs/reconstruction/kodim01_recon.png

Network Interpolation

If you want to use network interpolation, specify --interp_model and __interp_alpha.

    python codec.py decode -c CONFIG_PATH -m MODEL_PATH -i BIN_PATH -o SAVE_IMG_PATH --interp_model INTERP_MODEL_PATH --interp_alpha 0.8

The interpolated weight will be interp_alpha * torch.load(MODEL_PATH) + (1 - interp_alpha) * torch.load(INTERP_MODEL_PATH). You can balance the trade-off between distortion and perception by changing alpha.

Acknowledgments

We thank Jorge Pessoa for the code of GDN.