Readme

This work is published as paper in DCC 2022. Title: Rate Distortion Characteristic Modeling for Neural Image Compression Cite as: Jia C, Ge Z, Wang S, et al. Rate distortion characteristic modeling for neural image compression[C]//2022 Data Compression Conference (DCC). IEEE, 2022: 202-211.

Environment

First run

pip install -r requirements.txt

for the installazation of python packages.

Python Interface installation (you may encounter some bugs if use python2)

You need these files in the 'Util' folder：AE.cpp, My_Range_Coder.h, My_Range_Encoder.cpp, My_Range_Decoder.cpp.

Step. 1：

Change the Python.h direction in the AE.cpp file for the direction of Python.h in your system. (e.g. /usr/include/python3.6m/Python.h）

Step. 2：

python setup.py build
python setup.py install

Dataset

Training dataset includes 600K cropped 256*256 images generated from LIU4K dataset.

Reference: J. Liu, D. Liu, W. Yang, S. Xia, X. Zhang and Y. Dai, "A Comprehensive Benchmark for Single Image Compression Artifact Reduction," in IEEE Transactions on Image Processing, vol. 29, pp. 7845-7860, 2020, doi: 10.1109/TIP.2020.3007828.

Training procedure

The training of modnet adopts the weights of the highest rate fixed-rate model as initialization. Download the baseline_model.zip and unzip it into the 'baseline_model' folder and follow the instructions below to train modnet.

download url : https://drive.google.com/file/d/1qoOxiiRT_vQQfgN4v7epYaTH4CbXTfVu/view?usp=sharing

For mse model

In the /RDM4NIC/ directory run

python train.py --data * 

In * ,you can config as you need. (same below)

For msssim model

In the /RDM4NIC/ directory run

python train_msssim.py --data *

Testing procedure

Put the trained model into folder 'proposed_model' and run

python test.py --lmd * --input *' .