Lossless Image Compression through Super-Resolution

Sheng Cao, Chao-Yuan Wu, Philipp Krähenbühl.

[Paper]

Citation

@article{cao2020lossless,
  title={Lossless Image Compression through Super-Resolution},
  author={Cao, Sheng and Wu, Chao-Yuan and and Kr{\"a}henb{\"u}hl, Philipp},
  year={2020},
  journal={arXiv preprint arXiv:2004.02872},
}

If you use our codebase, please consider also citing L3C

Overview

This is the official implementation of SReC in PyTorch. SReC frames lossless compression as a super-resolution problem and applies neural networks to compress images. SReC can achieve state-of-the-art compression rates on large datasets with practical runtimes. Training, compression, and decompression are fully supported and open-sourced.

Getting Started

We recommend the following steps for getting started.

1. Install the necessary dependencies
2. Download the Open Images validation set
3. Run compression on Open Images validation set with trained model weights

Installation

See here for installation instructions.

Model Weights

We've released trained models for both ImageNet64 and Open Images (PNG). All compression results are measured in bits per subpixel (bpsp).

Training

To run code, you need to be in top level directory.

python3 -um src.train \
  --train-path "path to directory of training images" \
  --train-file "list of filenames of training images, one filename per line" \
  --eval-path "path to directory of eval images" \
  --eval-file "list of filenames of eval images, one filename per line" \
  --plot "directory to store model output" \
  --batch "batch size"

The training images must be organized in form of train-path/filename from filename in train-file. Same thing applies to eval images.

We've included our training and eval files used for ImageNet64 and Open Images (PNG) in datasets directory.

For ImageNet64, we use a slightly different set of hyperparameters than Open Images hyperparameters, which are the default. To train ImageNet64 based on settings from our paper, run

python3 -um src.train \
  --train-path "path to directory of training images" \
  --train-file "list of filenames of training images, one filename per line" \
  --eval-path "path to directory of eval images" \
  --eval-file "list of filenames of eval images, one filename per line" \
  --plot "directory to store model output" \
  --batch "batch size" \
  --epochs 10 \
  --lr-epochs 1 \
  --crop 64

Run python3 -um src.train --help for a list of tunable hyperparameters.

Evaluation

Given a model checkpoint, this evaluates theoretical bits/subpixel (bpsp) based on log-likelihood. The log-likelihood bpsp lower-bounds the actual compression bpsp.

python3 -um src.eval \
  --path "path to directory of images" \
  --file "list of filenames of images, one filename per line" \
  --load "path to model weights"

Compression/Decompression

With torchac installed, you can run compression/decompression to convert any image into .srec files. The following compresses a directory of images.

python3 -um src.encode \
  --path "path to directory of images" \ 
  --file "list of filenames of images, one filename per line" \
  --save-path "directory to save new .srec files" \
  --load "path to model weights"

If you want an accurate runtime, we recommend running python with -O flag to disable asserts. We also include an optional --decode flag so that you can check if decompressing the .srec file gives the original image, as well as provide runtime for decoding.

To convert .srec files into PNG, you can run

python3 -um src.decode \
  --path "path to directory of .srec images" \ 
  --file "list of filenames of .srec images, one filename per line" \
  --save-path "directory to save png files" \
  --load "path to model weights"

Downloading ImageNet64

You can download ImageNet64 training and validation sets here.

Preparing Open Images Dataset (PNG)

We use the same set of training and validation images of Open Images as L3C.

For validation images, you can download them here.

For training images, please clone the L3C repo and run script from here

See this issue for differences between Open Images JPEG and Open Images PNG.

Acknowledgment

Thanks to L3C for implementations of EDSR, logistic mixtures, and arithmetic coding. Special thanks to Fabian Mentzer for letting us know about issues with the preprocessing script for Open Images JPEG and resolving them quickly.