JND-LIC

Existing human visual perception-oriented image compression methods well maintain the perceptual quality of compressed images, but they may introduce fake details into the compressed images, and cannot dynamically improve the perceptual rate-distortion performance at the pixel level. To address these issues, a just noticeable difference (JND)-based learned image compression (JND-LIC) method is proposed for human visual perception in this paper, in which a weight-shared model is used to extract image features and JND features, and the learned JND features are utilized as perceptual prior knowledge to assist the image coding process. In order to generate a highly compact image feature representation, a JND-based feature transform module is proposed to model the pixel-to-pixel masking correlation between the image features and the JND features. Furthermore, inspired by eye movement research that the human visual system perceives image degradation unevenly, a JND-guided quantization mechanism is proposed for the entropy coding, which adjusts the quantization step of each pixel to further eliminate perceptual redundancies.

Dataset

The Flicker 2W dataset is used for training, which can be downloaded from this link. To get the corresponding JND maps, run the following script:

python Util.jnd.py

Traing and Testing

python train.py
python test.py

Requirements

• PyTorch
• Torchvision
• lpips
• numpy
• os
• compressai
• opencv-python

Citation

If you find our paper or code useful for your research, please cite:

@ARTICLE{10697103,
  author={Pan, Zhaoqing and Zhang, Guoyu and Peng, Bo and Lei, Jianjun and Xie, Haoran and Wang, Fu Lee and Ling, Nam},
  journal={IEEE Transactions on Broadcasting}, 
  title={JND-LIC: Learned Image Compression via Just Noticeable Difference for Human Visual Perception}, 
  year={2024},
  volume={},
  number={},
  pages={1-12},
  doi={10.1109/TBC.2024.3464413}}