LC-FDNet

[LC-FDNet : Learned Lossless Image Compression with Frequency Decomposition Network] CVPR 2022

Hochang Rhee, Yeong Il Jang, Seyun Kim, Nam Ik Cho

[Paper] [Supplementary] [Arxiv]

Environments

• Ubuntu 18.04
• Pytorch 1.7.0
• CUDA 10.0.130 & cuDNN 7.6.5
• Python 3.7.7

You can type the following command to easily build the environment. Download 'fdnet_env.yml' and type the following command.

conda env create -f fdnet_env.yml

Abstract

Recent learning-based lossless image compression methods encode an image in the unit of subimages and achieve better or comparable performances to conventional non-learning algorithms, along with reasonable inference time. However, these methods do not consider the performance drop in the high-frequency region, giving equal consideration to the low and high-frequency regions. In this paper, we propose a new lossless image compression method that proceeds the encoding in a coarse-to-fine manner to deal with low and high-frequency regions differently. We first compress the low-frequency components and then use them as additional input for encoding the remaining high-frequency region. The low-frequency components act as strong prior in this case, which leads to improved estimation in the high-frequency area. To classify the pixels as low or high-frequency regions, we also design a discriminating network that finds adaptive thresholds depending on the color channel, spatial location, and image characteristics. The network derives the image-specific optimal ratio of low/high-frequency components as a result. Experiments show that the proposed method achieves state-of-the-art performance for benchmark high-resolution datasets, outperforming both conventional learning-based and non-learning approaches.

Brief Description of Our Proposed Method

Framework of our compresion scheme

The framework of our compression scheme. Depending on the spatial location, each pixel is grouped as either $a,b,c,d$. The input image is split into subimages, which are sequentially compressed. The subimage $x_{YUV,a}$ is initially encoded using a conventional compression algorithm. The remaining subimages are compressed through deep networks, which receive the previously encoded subimages as input and compress the current subimage. The dotted arrow denotes that the corresponding subimage is currently being compressed. The compressed subimage is then used as an additional input for encoding the next subimage.

Architecture of LC-FDNet

The architecture of LC-FDNet. In this figure, we consider the case of compressing $y=x_{Y,d}$ given $x_{in}=x_{YUV,a}$. AFD part first receives $x_{in}$ and determines each pixel as belonging to either low or high-frequency regions, using error variance map $\sigma_y$ and error variance threshold $\tau_y$. Afterward, LFC encodes the low-frequency region of subimage $y$. HFC then receives the encoded low-frequency region as additional input and compresses the remaining high-frequency region. The decoding process is provided in the Supplementary Material.

Experimental Results

Comparison of our method with other non-learning and learning-based codes on high-resolution benchmark dataset. We measure the performances in bits per pixel (bpp). Best performance is highlighted in bold and the second-best performance is denoted with *. The difference in percentage to our method is highlighted in green.

Dataset

Train Dataset

[FLICKR2K] (https://github.com/limbee/NTIRE2017)

Test Dataset

[DIV2K] (https://data.vision.ee.ethz.ch/cvl/DIV2K/)

[CLIC] (http://clic.compression.cc/2019/challenge/)

Brief explanation of contents

|── experiments
    ├──> experiment_name 
         ├──> ckpt : trained models will be saved here
         └──> log  : log will be saved here
|── dataset
    ├──> dataset_name1 
         ├──> train : training images of dataset_name1 should be saved here
         └──> test  : test images of dataset_name1 should be saved here
    ├──> dataset_name2
         ├──> train : training images of dataset_name2 should be saved here
         └──> test  : test images of dataset_name2 should be saved here         
|── utils : files for utility functions
|── config.py : configuration should be controlled only here 
|── decode.py : decode compressed files to images
|── encode.py : encode images to compressed format
|── fdnet_env.yml : virtual enviornment specification
|── model.py : architecture of FDNet
|── test.py : test the model. performance is estimated (not actual compression)
|── train.py : train the model
└── jpegxl : folder for jpegxl library. explained below.

We use 'torchac' library as our arithmetic coder, which is available at https://github.com/fab-jul/torchac.

Guidelines for Codes

1. Install JPEG-XL

   1. Download JPEG-XL from https://gitlab.com/wg1/jpeg-xl and follow the installation process
   2. Change the directory name 'libjxl' to 'jpegxl'

2. Check configurations from config.py

3. Run the following command for training the network

python train.py --gpu_num=0 --experiment_name='default/' --dataset='div2k/'

The trained model will be saved in the following directory : experiments/default/ckpt

4. Run the following command for testing the network. Note that the estimated bpp is printed. The actual bpp differs from the estimated bpp (about 0.1 bpp)

   **** parameter empty_cache in config.py should be set to True if memory issue occurs ****

python test.py --gpu_num=0 --experiment_name='default/' --dataset='div2k/' --empty_cache=True

5. Run the following command for actual compression (encoding)

python encode.py --gpu_num=0 --experiment_name='default/' --dataset='div2k/' --empty_cache=True

The encoded result will be saved in 'encoded_results/'

6. Run the following command for decoding

python decode.py --gpu_num=0 --experiment_name='default/' --dataset='div2k/' --empty_cache=True

This decodes the compressed files in 'encoded_results/' folder.

The decoded result will be saved in 'decoded_results/'

Citation

If you use the work released here for your research, please cite this paper.

@InProceedings{Rhee_2022_CVPR,
    author    = {Rhee, Hochang and Jang, Yeong Il and Kim, Seyun and Cho, Nam Ik},
    title     = {LC-FDNet: Learned Lossless Image Compression With Frequency Decomposition Network},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {6033-6042}
}