Mitigating Long-tail Distribution in Oracle Bone Inscriptions: Dataset, Model, and Benchmark ☯️

The first attempt to apply diffusion model in realistic and controllable OBI generation

Jinhao Li^1*, Zijian Chen^2*, Runze Jiang³, Tingzhu Chen^3†, Changbo Wang^1†, Guangtao Zhai²

¹School of Computer Science and Technology, East China Normal University
²Institute of Image Communication and Information Processing, Shanghai Jiao Tong University
³School of Humanities, Shanghai Jiao Tong University
^*Both authors contributed equally to this research ^†Corresponding authors

Overview of the proposed Oracle-P15K dataset. The dataset comprises 14,542 OBI images with structure-aligned expert-annotated glyphs. Based on this, we present a pseudo OBI image generator, namely OBIDiff, to alleviate the long-tail distribution problem in current OBI datasets. Extensive experiments demonstrate both the necessity of Oracle-P15K and the effectiveness of OBIDiff in improving the performance of downstream OBI tasks.

Release 🚀

• [2025/4/13] ⚡️ Github repo for Oracle-P15K is online.

Motivations 💡

The existing OBI datasets suffer from a long-tail distribution problem. Consequently, OBI-related models achieve superior performance in majority classes while underperforming in minority classes. Therefore, we construct Oracle-P15K, a large-scale structure-aligned OBI dataset comprising 14,542 images infused with domain knowledge from OBI experts. The Oracle-P15K dataset can also serve as a comprehensive benchmark for researchers to develop and evaluate their methods for dealing with other OBI information processing tasks, such as OBI denoising, recognition, etc.

Construction Pipeline 🧩

Focusing on structure-aligned image pairs for OBI generation and denoising models.

Pseudo OBI Generator 🤖

Our OBIDiff consists of an autoencoder, a stable diffusion (SD) model, a glyph encoder, and a style encoder. Given a clean glyph image and a target rubbing-style image, it can effectively transfer the noise style of the original rubbing to the glyph image.

Results on OBI Generation and Denoising Tasks 📌

Qualitative results on the OBI generation tasks (click to expand)

Quantitative results on the OBI generation tasks (click to expand)

• Fitted kernel distribution of four low-level features including brightness, contrast, sharpness, and spatial information (SI):

• Recognition accuracy of augmented images generated by the proposed OBIDiff and other OBI generation methods w.r.t. the scale of data augmentation:

Qualitative results on the OBI denoising tasks (click to expand)

Quantitative results on the OBI denoising tasks (click to expand)

User Preference Study 👥

We develop a web-based user interface with automated navigation to facilitate the evaluation process.

Contact ✉️

Please contact the first author of this paper for queries.

• Jinhao Li, [email protected]

Citation 📎

If you find our work interesting, please feel free to cite our paper:

@misc{li2025mitigatinglongtaildistributionoracle,
      title={Mitigating Long-tail Distribution in Oracle Bone Inscriptions: Dataset, Model, and Benchmark}, 
      author={Jinhao Li and Zijian Chen and Runze Dong and Tingzhu Chen and Changbo Wang and Guangtao Zhai},
      year={2025},
      eprint={2504.09555},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.09555}, 
}

Acknowledgements 🏆

This work was supported by the National Social Science Foundation of China (24Z300404220) and the Shanghai Philosophy and Social Science Planning Project (2023BYY003).