We build a real-world degraded character image dataset by selecting from the historical Chinese character and oracle document datasets. The reason for selecting such images is that most of them contain complex real-world degradation. The dataset includes training and testing sets consisting of noisy-clean character image pairs.
Where:
- denoise: The benchmark for character image denoising, which includes real-world noise. We produce these images by removing noise from the original character images.
- restore: The benchmark for character image restoration, which includes real-world noise. We produce these images by removing noise from the original character images and also repairing the broken characters.
- gray: The benchmark for character image denoising, which includes real-world noise with a gray background. We produce these images by removing noise from the original character images.
- test: data for testing purposes.
For any kind of use of these datasets, please cite the following:
@inproceedings{shi2022rcrn,
title={RCRN: Real-world character image restoration network via skeleton extraction},
author={Shi, Daqian and Diao, Xiaolei and Tang, Hao and Li, Xiaomin and Xing, Hao and Xu, Hao},
booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
pages={1177--1185},
year={2022}
}