A Fast Attentive Denoising Framework for Oracle Bone Inscriptions
2Institute of Image Communication and Information Processing, Shanghai Jiao Tong University
3School of Humanities, Shanghai Jiao Tong University
4Center for the Study and Application of Chinese Characters, East China Normal University
†Corresponding author
The overall architecture of our OBIFormer. (a) OBIFormer block (OFB) that injects glyph information into the denoising backbone, (b) Glyph structural network block (GSNB) that extracts glyph features, (c) Channel-wise self-attention block (CSAB) that generates channel-wise self-attention effectively and efficiently, (d) Selective kernel feature fusion (SKFF) module that aggregates reconstruction features and glyph features.
- [2025/6/17] ⚡️ Pre-trained models are released !
- [2025/6/17] ⚡️ Github repo for OBIFormer is online !
Train the model from the scratch:
python train.py --train_input /path/to/input/of/train --train_target /path/to/target/of/train --val_input /path/to/input/of/val --val_target /path/to/target/of/val --store_path /path/to/results
Test a trained model:
python test.py --input_path /path/to/input --store_path /path/to/results --checkpoint /path/to/checkpoint
Checkpoints are available at the pretrained folder.
The previous OBI denoising methods either focus on pixel-level information or utilize vanilla transformers for glyph-based OBI denoising, which leads to tremendous computational overhead. Therefore, we proposed a fast attentive denoising framework for oracle bone inscriptions, i.e., OBIFormer. It leverages channel-wise self-attention, glyph extraction, and selective kernel feature fusion to reconstruct denoised images precisely while being computationally efficient. Our OBIFormer achieves state-of-the-art denoising performance for PSNR and SSIM metrics on synthetic and original OBI datasets. Furthermore, comprehensive experiments on a real oracle dataset demonstrate the great potential of our OBIFormer in assisting automatic OBI recognition.
We summarize and examplify the previous OBI datasets.
Summary
Examples
Quantitative results on Oracle-50K and NCIB datasets (click to expand)
Qualitative results on Oracle-50K dataset (click to expand)
Qualitative results on NCIB dataset (click to expand)
To further validate the effectiveness of OBI denoising in improving recognition accuracy, we employed ResNet-18,ResNet-50,and ResNet-152 for the OBI recognition task on the test set of the Oracle-50K dataset.
We evaluate the number of parameters (#Param.), FLOPs, and infer time of different OBI denoising methods.
To explore the generalization ability of our OBIFormer, we tested it on a real oracle dataset (i.e., the OBC306 dataset) after training it on Oracle-50K and NCIB datasets.
Qualitative results on NCIB dataset (click to expand)
Qualitative results on NCIB dataset (click to expand)
We visualize the deep features extracted by the last OFB, which consists of reconstruction and glyph features.
Please contact the first author of this paper for queries.
- Jinhao Li,
lomljhoax@stu.ecnu.edu.cn
If you find our work interesting, please feel free to cite our paper:
@misc{li2025obiformerfastattentivedenoising,
title={OBIFormer: A Fast Attentive Denoising Framework for Oracle Bone Inscriptions},
author={Jinhao Li and Zijian Chen and Tingzhu Chen and Zhiji Liu and Changbo Wang},
year={2025},
eprint={2504.13524},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2504.13524},
}
This work was supported by the National Social Science Foundation of China (24Z300404220) and the Shanghai Philosophy and Social Science Planning Project (2023BYY003).