SHOMEI: Per-Character Certified Robustness for Japanese OCR via Kanji Confusion-Pool Smoothing

Barbhaya, Mayur

doi:10.51094/jxiv.5902

##article.authors##

Barbhaya, Mayur Research and Development Department, VeBuIn Pvt. Ltd.

DOI:

https://doi.org/10.51094/jxiv.5902

キーワード:

Japanese OCR、 certified robustness、 randomised smoothing、 confusion pool、 kanji、 document AI、 Levine-Feizi smoothing、 Per-Character Certified Robustness for Japanese OCR

抄録

Japanese OCR systems, including recent vision-language models, produce per-character output with only heuristic confidence indicators such as softmax probability. In high-stakes documents — legal contracts, medical prescriptions, government notices — a silent substitution of a visually similar kanji (末 for 未, 士 for 土, 己/已/巳) can invert clause meaning, yet no OCR system in current use provides a mathematically grounded per-character guarantee that its reading is stable under such substitutions. I propose SHOMEI (証明, meaning proof in Japanese), a model-agnostic wrapper that equips any pretrained Japanese OCR with per-character formal certificates against a discrete image-side glyph-substitution adversary drawn from a language-derived confusion pool. The framework applies the Levine–Feizi discrete-smoothing bound over per-position substitutions, and emits either a character together with an integer certified radius, or an [ABSTAIN] token when no such certificate can be issued. A 27,531-character four-tier Japanese confusion pool combining empirical similarity (Yencken), radical-structural neighbours (CHISE IDS), Unicode CJK-Unified cross-script duplicates, and curated kana/digit lookalikes is released as a companion asset under CC-BY 4.0. I evaluate SHOMEI across three regimes: clean synthetic renders (84% coverage at 100% certification accuracy), degraded synthetic input (56% at 100%), and 11 real-world Japanese scanned documents (34% coverage at 96% certification accuracy overall, 100% on modern receipts). A base-OCR ablation across Tesseract, Azure Document Intelligence and Azure OpenAI GPT-5.1 reveals that SHOMEI's coverage is governed by base context-override behaviour rather than raw pixel accuracy. Based on a review of the prior literature, this is the first per-character certified-robustness framework for Japanese OCR.

利益相反に関する開示

The author declares no conflicts of interest in relation to this work.

ダウンロード *前日までの集計結果を表示します

ダウンロード実績データは、公開の翌日以降に作成されます。

引用文献

Cohen, J., Rosenfeld, E., and Kolter, J. Z. (2019). Certified adversarial robustness via randomized smoothing. In International Conference on Machine Learning (ICML), pp. 1310-1320.

Levine, A., and Feizi, S. (2020). Robustness certificates for sparse adversarial attacks by randomized ablation. In AAAI Conference on Artificial Intelligence, pp. 4585-4593.

Ye, M., Gong, C., and Liu, Q. (2020). SAFER: A structure-free approach for certified robustness to adversarial word substitutions. In Annual Meeting of the Association for Computational Linguistics (ACL), pp. 3465-3475.

Zhang, X., Hong, H., Hong, Y., Huang, P., Wang, B., Ba, Z., and Ren, K. (2024). Text-CRS: A generalized certified robustness framework against textual adversarial attacks. IEEE Symposium on Security and Privacy.

Li, L., Weber, M., Xu, X., Rimanic, L., Kailkhura, B., Xie, T., Zhang, C., and Li, B. (2022). TSS: Transformation-specific smoothing for robustness certification. In ACM CCS.

Salman, H., Jain, S., Wong, E., and Madry, A. (2021). Certified patch robustness via smoothed vision transformers. arXiv:2110.07719.

Fujitake, M. (2024). JaPOC: Japanese post-OCR correction benchmark using vouchers. In PRICAI 2024, Springer, pp. 45-58.

Huang, Z., Erdogdu, M. A., et al. (2023). RS-Del: Edit distance robustness certificates for sequence classifiers via randomized deletion. arXiv:2302.01757.

Shao, D., Fesser, L., and Kwiatkowska, M. (2024). STR-Cert: Robustness certification for deep text recognition on deep learning pipelines and vision transformers. arXiv:2401.05338.

Cui, C., Sun, T., Liang, S., Gao, T., Zhang, Z., Liu, J., Wang, X., Zhou, C., Liu, H., Lin, M., et al. (2025). PaddleOCR-VL: Boosting multilingual document parsing via a 0.9B ultra-compact vision-language model. arXiv:2510.14528.

Wei, H., Liu, C., Chen, J., Wang, J., Kong, L., Xu, Y., Ge, Z., Zhao, L., Sun, J., Peng, Y., Han, C., and Zhang, X. (2024). General OCR Theory: Towards OCR-2.0 via a unified end-to-end model. arXiv:2409.01704.

Yencken, L. (2010). Kanji-confusion dataset. https://lars.yencken.org/datasets/kanji-confusion/, CC-BY 3.0.

CHISE Project. CHISE IDS database (Ideographic Description Sequences for CJK-Unified characters). https://github.com/cjkvi/cjkvi-ids.

Wei, H., Sun, Y., and Li, Y. (2025). DeepSeek-OCR: Contexts optical compression. arXiv:2510.18234.

Baek, J., Egashira, K., Onohara, S., Miyai, A., Imajuku, Y., Ikuta, H., and Aizawa, K. (2025). MangaVQA and MangaLMM: A benchmark and specialized model for multimodal manga understanding. arXiv:2505.20298.

Sasagawa, K., Kurita, S., and Kawahara, D. (2025). Evaluating multimodal large language models on vertically written Japanese text. arXiv:2511.15059.

Aizawa, K., Fujimoto, A., Otsubo, A., Ogawa, T., Matsui, Y., Tsubota, K., and Ikuta, H. (2020). Building a manga dataset "Manga109" with annotations for multimedia applications. IEEE MultiMedia, 27(2), 8-18.

Yang, X., Arora, A., Jheng, S., and Dell, M. (2023). Quantifying character similarity with vision transformers. EMNLP 2023, pp. 14185-14202. arXiv:2305.14672.

Cheng, X., Xu, W., Chen, K., Jiang, S., Wang, F., Wang, T., Chu, W., and Qi, Y. (2020). SpellGCN: Incorporating phonological and visual similarities into language models for Chinese spelling check. ACL, pp. 871-881. arXiv:2004.14166.

Xu, H., Li, Z., Zhou, Q., Li, C., Wang, Z., Cao, Y., Huang, H., and Mao, X. (2021). Read, listen, and see: Leveraging multimodal information helps Chinese spell checking. Findings of ACL, pp. 716-728. arXiv:2105.12306.