适用于内容安全审核的图像文本识别质量评估方法

张玉龙; 梁天一; 张茹; 刘功申

doi:10.11959/j.issn.2096-109x.AQ25222

您当前的位置：

首页 >

文章列表页 >

适用于内容安全审核的图像文本识别质量评估方法

学术论文 | 更新时间：2026-05-11

- 适用于内容安全审核的图像文本识别质量评估方法
- Quality assessment method for image text recognition in content security review
- 网络与信息安全学报 2026年12卷第2期页码：55-64
- 作者机构：
  
  1.北京邮电大学网络空间安全学院，北京 100876
  2.上海交通大学计算机学院（网络空间安全学院），上海 200240
  3.华东师范大学计算机科学与技术学院，上海 200062
- 作者简介：
  
  [ "张玉龙（1999- ），男，北京邮电大学硕士生，主要研究方向为多媒体信息内容安全、人工智能安全、大模型。" ]
  [ "梁天一（1998- ），男，华东师范大学博士生，主要研究方向为多模态生成式模型。" ]
  [ "张茹（1976- ），女，博士，北京邮电大学教授、博士生导师，主要研究方向为信息隐藏、人工智能安全。" ]
  [ "刘功申（1974- ），男，博士，上海交通大学教授、博士生导师，主要研究方向为人工智能安全、自然语言处理、信息安全。" ]
- 基金信息：
  
  国家自然科学基金;The National Natural Science Foundation of China(U21B2020)
- DOI：10.11959/j.issn.2096-109x.AQ25222
  中图分类号： TP393
- 收稿：2025-08-24，
  
  修回：2026-04-14，
  
  录用：2026-04-15，
  
  纸质出版：2026-04-25
- 稿件说明：
移动端阅览
张玉龙, 梁天一, 张茹, 等. 适用于内容安全审核的图像文本识别质量评估方法[J]. 网络与信息安全学报, 2026, 12(2): 55-64.

Zhang Yulong, Liang Tianyi, Zhang Ru, et al. Quality assessment method for image text recognition in content security review[J]. Chinese Journal of Network and Information Security, 2026, 12(2): 55-64.
张玉龙, 梁天一, 张茹, 等. 适用于内容安全审核的图像文本识别质量评估方法[J]. 网络与信息安全学报, 2026, 12(2): 55-64. DOI： 10.11959/j.issn.2096-109x.AQ25222.

Zhang Yulong, Liang Tianyi, Zhang Ru, et al. Quality assessment method for image text recognition in content security review[J]. Chinese Journal of Network and Information Security, 2026, 12(2): 55-64. DOI： 10.11959/j.issn.2096-109x.AQ25222.

摘要

针对内容安全审核场景下图像文本识别质量评估的难题，提出基于一致性置信（CC）机制的质量评估方法。现有评估方法存在显著不足，传统方法依赖人工标注且成本高昂，视觉语言模型作为评判器（VLM-as-Judge）方法存在逻辑悖论和偏见，不确定性量化方法需要访问模型内部参数。基于CC机制的质量评估方法通过多模型一致性分析实现无监督质量评估，利用编辑距离量化一致性程度，并设计动态权重和阈值门控机制识别低质量输出。该方法具有黑盒评估、完全无监督、模型无关性三大核心优势。实验结果表明：在两个数据集上质量验证F1得分达49.9%~51.61%，相比基线方法提升42.76%；在对抗性文本、复杂场景等挑战性场景下均表现最佳；模型集成使低成本开源模型超越高成本闭源模型，性能提升达32分。

Abstract

A quality assessment method based on the consistency confidence (CC) mechanism was proposed to address the challenge of image text recognition quality assessment in content security auditing scenarios. Significant deficiencies were identified in existing evaluation methods. Traditional methods were found to rely on manual annotation and be cost-prohibitive. The visual language model as judge (VLM-as-Judge) approach was observed to suffer from logical paradoxes and biases. Uncertainty quantification methods were shown to require access to internal model parameters. With the proposed CC-based method

unsupervised quality assessment was achieved through multi-model consistency analysis. Edit distance was utilized to quantify the degree of consistency

and dynamic weighting with threshold gating mechanisms was designed to identify low-quality outputs. This method possesses three major advantages: black-box evaluation

complete unsupervisedness

and model-agnostic property. Experimental results demonstrated that the quality verification F1 score reached 49.9%~51.61%

which represented a 42.76% improvement over baseline methods. The best performance was exhibited in challenging scenarios such as adversarial text and complex scenes. Through model ensemble

low-cost open-source models were enabled to surpass high-cost closed-source models

with a performance improvement of 32 points.

关键词

Keywords

references

刘帆 , 王凤美 . 多模态内容安全审核系统构建思考 [J ] . 中国传媒科技 , 2023 ( 4 ): 149 - 153 .

Liu F , Wang F M . Thoughts on the construction of multimodal content security audit system [J ] . Media Science and Technology of China , 2023 ( 4 ): 149 - 153 .

乔喆 . 人工智能生成内容技术在内容安全治理领域的风险和对策 [J ] . 电信科学 , 2023 , 39 ( 10 ): 136 - 146 .

Qiao Z . Risks and countermeasures of artificial intelligence generated content technology in content security governance [J ] . Telecommunications Science , 2023 , 39 ( 10 ): 136 - 146 .

李亚玲 , 覃缘琪 , 魏阙 . 人工智能生成内容的潜在风险及治理对策 [J ] . 智能科学与技术学报 , 2023 , 5 ( 3 ): 415 - 423 .

Li Y L , Qin Y Q , Wei Q . Potential risks and governance strategies of artificial intelligence generated content technology [J ] . Chinese Journal of Intelligent Science and Technology , 2023 , 5 ( 3 ): 415 - 423 .

王润民 , 桑农 , 丁丁 , 等 . 自然场景图像中的文本检测综述 [J ] . 自动化学报 , 2018 , 44 ( 12 ): 2113 - 2141 .

Wang R M , Sang N , Ding D , et al . Text detection in natural scene image: a survey [J ] . Acta Automatica Sinica , 2018 , 44 ( 12 ): 2113 - 2141 .

Zhang D Z , Yu Y H , Dong J H , et al . MM-LLMs: recent advances in MultiModal large language models [C ] // Proceedings of the Findings of the Association for Computational Linguistics ACL 2024 . Stroudsburg : ACL , 2024 : 12401 - 12430 .

Duan H D , Yang J M , Qiao Y X , et al . VLMEvalKit: an open-source ToolKit for evaluating large multi-modality models [C ] // Proceedings of the 32nd ACM International Conference on Multimedia . New York : ACM Press , 2024 : 11198 - 11201 .

Liu Y L , Li Z , Huang M X , et al . OCRBench: on the hidden mystery of OCR in large multimodal models [J ] . Science China Information Sciences , 2024 , 67 ( 12 ): 220102 .

Poznanski J , Rangapur A , Borchardt J , et al . olmOCR: unlocking trillions of tokens in PDFs with vision language models [PP ] . arXiv ( 2025-07-02 )[ 2025-08-24 ] . arXiv: arXiv. 2502 . 18443 .

Zhang R Y , Zhang H , Zheng Z D . VL-uncertainty: detecting hallucination in large vision-language model via uncertainty estimation [PP ] . arXiv ( 2024-11-28 )[ 2025-08-24 ] . arXiv: arXiv. 2411.11919.

Liu X O , Chen T J , Da L C , et al . Uncertainty quantification and confidence calibration in large language models: a survey [C ] // Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 . New York : ACM Press , 2025 : 6107 - 6117 .

Blecher L , Cucurull G , Scialom T , et al . Nougat: neural optical understanding for academic documents [PP ] . V1 . ( 2023-08-25 )[ 2025-08-24 ] . arXiv: arXiv. 2308 . 13418 .

Wang B , Xu C , Zhao X M , et al . MinerU: an open-source solution for precise document content extraction [PP ] . V1 . ( 2024-09-27 )[ 2025-08-24 ] . arXiv: arXiv. 2409 . 18839 .

Shen Z J , Lo K , Wang L L , et al . VILA: improving structured content extraction from scientific PDFs using visual layout groups [J ] . Transactions of the Association for Computational Linguistics , 2022 , 10 : 376 - 392 .

Huang Y P , Lyu T C , Cui L , et al . LayoutLMv3: pre-training for document AI with unified text and image masking [C ] // Proceedings of the 30th ACM International Conference on Multimedia . New York : ACM Press , 2022 : 4083 - 4091 .

Zhong X , Tang J B , Jimeno Yepes A . PubLa y Net: largest dataset ever for document layout analysis [C ] // Proceedings of the 2019 International Conference on Document Analysis and Recognition (ICDAR) . Piscataway : IEEE Press , 2019 : 1015 - 1022 .

Liu C L , Wei H R , Chen J Y , et al . Focus anywhere for fine-grained multi-page document understanding [PP ] . arXiv ( 2024-05-23 )[ 2025-08-24 ] . arXiv: arXiv. 2405 . 14295 .

Chiang W L , Gonzalez J , Li D C , et al . Judging LLM-as-a-judge with MT-bench and chatbot arena [C ] // Proceedings of the Advances in Neural Information Processing Systems 36 . California : Neural Information Processing Systems Foundation, Inc. , 2023 : 46595 - 46623 .

Zhu L H , Wang X G , Wang X L . JudgeLM: fine-tuned large language models are scalable judges [PP ] . arXiv ( 2023-10-26 )[ 2025-08-24 ] . arXiv: arXiv. 2310 . 17631 .

Gu J W , Jiang X H , Shi Z C , et al . A survey on LLM-as-a-judge [J ] . The Innovation , 2026 : 101253 .

Chang Y P , Wang X , Wang J D , et al . A survey on evaluation of large language models [J ] . ACM Transactions on Intelligent Systems and Technology , 2024 , 15 ( 3 ): 1 - 45 .

Rando J , Zhang J , Carlini N , et al . Adversarial ML problems are getting harder to solve and to evaluate [PP ] . arXiv ( 2025-02-04 )[ 2025-08-24 ] . arXiv: arXiv. 2502.02260.

Albladi A , Islam M , Das A , et al . Hate speech detection using large language models: a comprehensive review [J ] . IEEE Access , 2025 , 13 : 20871 - 20892 .

Lu X Y , Zhang T K , Meng C , et al . VLM as policy: common-law content moderation framework for short video platform [C ] // Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 . New York : ACM Press , 2025 : 4682 - 4693 .

Wang X Z , Wei J , Schuurmans D , et al . Self-consistency improves chain of thought reasoning in language models [PP ] . arXiv ( 2022-03-21 )[ 2025-08-24 ] . arXiv: arXiv. 2203 . 11171 .

Li Y J , Liu B . A normalized levenshtein distance metric [J ] . IEEE Transactions on Pattern Analysis and Machine Intelligence , 2007 , 29 ( 6 ): 1091 - 1095 .

浏览量

下载量

CSCD

文章被引用时，请邮件提醒。

提交

工具集

关联资源

暂无数据