fl0wbar / EIC-Balle2017

Experiments of applying "End-to-end optimized image compression" Balle, 2017 on CLIC-2019 dataset

EIC(Balle 2017)

Experiments done in January 2019

The repo is a re-implementation of the model published in:

"End-to-end optimized image compression"
J. Ballé, V. Laparra, E. P. Simoncelli
https://arxiv.org/abs/1611.01704

Code for reproducing the results are modified from Tensorflow Compression (tensorflow 1.13 compatible commit)

This directory contains Tensorflow Compression code modified to use range coding ops in the tensorflow.contrib.coder.python.ops in tensorflow 1.13

Do not use with original Tensorflow Compression as it has fast custom kernels for range coding.

This was done to avoid compilation issues.

Please refer to README from Tensorflow > 2.0 compatible Tensorflow Compression (https://github.com/tensorflow/compression) repo for better code for reproducing these results.

This repo is a public backup of a hobby project.

CLIC Dataset

For script options

$ python eeicballe17.py -h

Training

$ python eeicballe17.py -v --train_glob="./datasets/CLIC/professional/train/*.png" --checkpoint_dir=./models/balle17CLIC train

Compressing and Decompressing single image using trained model

• Compress using model trained on CLIC 2019

$ python eeicballe17.py --verbose --checkpoint_dir=./models/balle17CLIC compress ./tests/groundtruth/kodak/kodim01.png ./tests/reconCLIC/epoch30684/kodim01.tfic

• Compress using model trained on BSDS500

$ python eeicballe17.py --verbose --checkpoint_dir=./models/balle17BSDS500 compress ./tests/groundtruth/kodak/kodim01.png ./tests/reconBSDS500/kodim01_bsds.tfic

• Decompress using model trained on CLIC 2019

$ python eeicballe17.py --verbose --checkpoint_dir=./models/balle17CLIC decompress ./tests/reconCLIC/epoch30684/kodim01.tfic ./tests/reconCLIC/epoch30684/kodim01_recon.png

• Decompress using model trained on BSDS500

$ python eeicballe17.py --verbose --checkpoint_dir=./models/balle17BSDS500 decompress ./tests/reconBSDS500/kodim01_bsds.tfic ./tests/reconBSDS500/kodim01_recon.png

Results on CLIC2019 dataset

Example

Original	EIC

Original: 491.1 kB
EIC:  Compressed File size: 15.1 kB
      Mean squared error: 29.7069
      PSNR (dB): 33.40
      Multiscale SSIM: 0.9706
      Multiscale SSIM (dB): 15.32
      Information content in bpp: 0.3120
      Actual bits per pixel: 0.3148

The image shown is an out-of-sample instance from the Kodak dataset. The EIC image is obtained by reconstruction via a learned model on CLIC-2019.

Note that the learned model was not adapted in any way for evaluation on this image.

bpp (bits-per-pixel)

mse (mean-squared-error)

loss (mse + entropy coding auxiliary loss)

probabilty mass function plot of entropy coder

tensorboard model graph

Citation

"End-to-end optimized image compression"
J. Ballé, V. Laparra, E. P. Simoncelli
https://arxiv.org/abs/1611.01704

"Efficient nonlinear transforms for lossy image compression"
J. Ballé
https://arxiv.org/abs/1802.00847

If you use this library for research purposes, please cite:

@software{tfc_github,
  author = "Ballé, Johannes and Hwang, Sung Jin and Johnston, Nick",
  title = "{T}ensor{F}low {C}ompression: Learned Data Compression",
  url = "http://github.com/tensorflow/compression",
  version = "1.2b1 (beta)",
  year = "2018",
}

In the above BibTeX entry, names are top contributors sorted by number of commits. Please adjust version number and year according to the version that was actually used.