tgjjj / EARN

Efficient Artifact Reduction Network for JPEG images

EARN

The source code of paper "EARN: toward efficient and robust JPEG compression artifact reduction", https://link.springer.com/article/10.1007/s00371-023-03008-4

Results

As shown below, EARN can effectively reduce the artifacts generated by JPEG compression and achieved state-of-the-art restoration performance with less computational costs and network parameters.

JPEG Quality=10

EARN Restored

Performance vs Parameters and Computation

Guide

Environment

For preparing running environment, run

conda env create -f EARN_env.yaml

Inference/Test

For inferencing or testing, run

python Inference_Metric_RGB.py

or

python Inference_Metric_Y.py

depending on the image type (color or gray). Demonstration inference results are shown in "./results" folder and our pretrained models are provided as "model_RGB.pth" and "model_Y.pth".

Train

For training, first construct your own training dataset. Please refer to "./LIVE1" and "./LIVE1_Y" folders for scripts and demonstrations about the how to construct it. Then change the dataset path in "Train_RGB.py" or "Train_Y.py" and run

python Train_RGB.py

or

python Train_Y.py

according to your dataset type (color or gray).

References

Please cite our paper if the repository helps you.

Teng, G., Jiang, R., Liu, X. et al. EARN: toward efficient and robust JPEG compression artifact reduction. Vis Comput (2023). https://doi.org/10.1007/s00371-023-03008-4