AI News
开源项目 GitHub Topic rag 2026-07-06T12:02 正文 ok 1306 words

📦 PaddlePaddle/PaddleOCR：将任意PDF或图片转为LLM可用的结构化数据

英文标题：📦 PaddlePaddle/PaddleOCR

PaddleOCR 是一个轻量级、高精度的OCR工具包，支持100多种语言，可将PDF和图片转换为JSON或Markdown格式的结构化数据。最新版PP-OCRv6在检测与识别准确率上分别提升4.6%和5.1%，统一模型覆盖50种语言，推理速度在CPU上提升5.2倍。该工具已深度集成Dify、RAGFlow、Cherry Studio等主流AI Agent项目，成为构建智能RAG和Agent应用的基础设施。
打开原文相关 2.8 · 机会 3
media448.3 KB media1.2 KB media1.3 KB media1.3 KB media1.3 KB media1.3 KB media3.9 KB media1.9 KB media1.3 KB media530.6 KB media3.0 MB media2.9 MB media5.1 MB media2.8 MB media8.2 KB media63.3 KB media2.7 KB media2.7 KB media2.7 KB media2.7 KB

中文

PaddleOCR 将 PDF 文档和图像以业界领先的准确率转换为结构化、可直接供大语言模型（LLM）使用的数据（JSON/Markdown）。凭借超过7万 Star 的社区认可以及被 Dify、RAGFlow、Cherry Studio 等顶级项目集成，PaddleOCR 是构建智能 RAG 和 Agentic 应用的基石。

📄 智能文档解析（为LLM就绪）

将杂乱的视觉信息转化为结构化数据，迎接大语言模型时代。

  • SOTA 文档 VLM：搭载 PaddleOCR-VL-1.6 (0.9B)，业界领先的轻量级视觉语言模型，专为文档解析设计。在 OmniDocBench v1.6 上达到 96.3% 的准确率，在文本、公式和表格识别方面领先，并显著增强了对古籍、生僻字、印章和图表的处理能力，以 Markdown 和 JSON 格式输出结构化结果。
  • 结构感知转换：由 PP-StructureV3 驱动，将复杂 PDF 和图像无缝转换为 Markdown 或 JSON。与 PaddleOCR-VL 系列模型不同，它提供更细粒度的坐标信息，包括表格单元格坐标、文本坐标等。
  • 生产级效率：以极小的模型体积实现商业级精度。在公开基准测试中超越众多闭源方案，同时保持资源高效，适合边缘/云端部署。

🔍 通用文字识别（场景OCR）

高速、多语种文字检测与识别的全球黄金标准。

  • 支持100+语言：原生支持庞大的全球语言库。PP-OCRv6 使用单一统一模型支持 50 种语言（中文、英文、日文及46种拉丁字母书写的语言），处理多语言文档时无需切换模型。
  • 复杂元素精通：不仅限于标准文字识别，我们支持在多种环境下的自然场景文字检测与识别，涵盖身份证、街景、书籍和工业部件等。
  • 性能飞跃：PP-OCRv6 相比 PP-OCRv5 检测准确率提高 4.6%，识别准确率提高 5.1%，超越主流视觉语言模型。端到端 CPU 推理速度提升 5.2 倍。

🛠️ 面向开发者生态

  • 无缝集成：AI Agent 生态的首选，已与 Dify、RAGFlow、Pathway 和 Cherry Studio 深度集成。
  • LLM 数据飞轮：提供构建高质量数据集的完整流水线，为微调大型语言模型提供可持续的“数据引擎”。
  • 一键部署：支持多种硬件后端（NVIDIA GPU、Intel CPU、昆仑芯 XPU 及各类 AI 加速器）。

🔥 2026.06.11：发布 PaddleOCR 3.7.0

  • PP-OCRv6 亮点：

    • 精度提升：中尺寸模型相比 PP-OCRv5_server 在检测上提高 4.6%，识别上提高 5.1%，超越主流 VLM（如 Qwen3-VL-235B、GPT-5.5），参数量仅 34.5M。
    • 50 种语言统一：单一模型覆盖中文、英文、日文及 46 种拉丁字母书写的语言，无需切换模型。
    • 专用场景：在数字显示屏、点阵字符、轮胎印痕和工业文本识别方面实现重大改进。
    • 推理加速：CPU 加速 5.2 倍（OpenVINO），Apple M4 上加速 6.1 倍（tiny），A100 GPU 上仅需 0.13 秒。
    • 全场景三级体系：tiny (1.5M) / small (7.7M) / medium (34.5M)，分别用于边缘、移动端和服务器部署。
    • 模型获取：所有模型均可在 HuggingFace 和 ModelScope 下载。

2026.05.28：发布 PaddleOCR 3.6.0 2026.04.21：发布 PaddleOCR 3.5.0

  • 灵活的推理后端：在 Paddle 静态图、Paddle 动态图或 Transformers 之间无缝切换。PaddleOCR 现已与 Hugging Face 生态深度集成，20 个主要模型支持以 Transformers 作为推理后端。
  • 办公文档转 Markdown：将 Word、Excel、PowerPoint 等常见文档格式转换为 Markdown。
  • 解析结果导出 DOCX：PaddleOCR-VL 系列、PP-StructureV3 以及 PP-DocTranslation 现已支持将解析结果导出为 DOCX，方便在 Microsoft Word 中查看和编辑。
  • 官方浏览器推理 SDK：发布 PaddleOCR.js，官方浏览器推理 SDK，支持在浏览器中直接运行 PP-OCRv5。

2026.01.29：发布 PaddleOCR 3.4.0

  • PaddleOCR-VL-1.5（SOTA 0.9B VLM）：我们最新的文档解析旗舰模型正式上线！
    • OmniDocBench 上 94.5% 准确率：超越顶级通用大模型和专业文档解析器。
    • 真实场景鲁棒性：率先引入 PP-DocLayoutV3 算法，实现不规则形状定位，攻克 5 大棘手场景：倾斜、扭曲、扫描件、光照影响和屏幕拍摄。
    • 能力拓展：现已支持印章识别、文字检测识别，并将语言种类扩展至 111 种（包括中国的藏文和孟加拉文）。
    • 长文档处理：支持自动跨页表格合并和层级标题识别。
    • 立即体验：可在 HuggingFace 或我们的官方网站获取。

2025.10.16：发布 PaddleOCR 3.3.0 2025.08.21：发布 PaddleOCR 3.2.0

更新日志

PaddleOCR 官网提供交互式体验中心和 API——无需搭建环境，一键体验。

👉 访问官方网站

如需本地使用，请根据需求参考以下文档：

🔄 执行结果快速预览

⭐ 收藏此仓库，关注强大的 OCR 和文档解析能力的最新动态与发布！ ⭐

👩‍👩‍👧‍👦 社区

PaddlePaddle 微信公众号 加入技术讨论群

😃 使用 PaddleOCR 的优秀项目

PaddleOCR 能走到今天，离不开出色的社区！💗 衷心感谢所有长期合作的伙伴、新的协作者以及每一位为 PaddleOCR 倾注热情的朋友——无论是否提及名字，你们的支持是我们前进的动力！

项目名称 描述
Dify 生产级 Agentic 工作流开发平台。
RAGFlow 基于深度文档理解的 RAG 引擎。
pathway Python ETL 框架，用于流处理、实时分析、LLM 流水线和 RAG。
MinerU 多类型文档转 Markdown 转换工具。
Umi-OCR 免费、开源、离线的批量 OCR 软件。
cherry-studio 支持多个 LLM 提供商的桌面客户端。
haystack AI 编排框架，用于构建可定制、生产级的 LLM 应用。
OmniParser OmniParser：面向纯视觉 GUI Agent 的屏幕解析工具。
QAnything 基于任何内容的问答系统。
了解更多项目 基于 PaddleOCR 的更多项目

本项目基于 Apache 2.0 许可证发布。

@misc{cui2025paddleocr30technicalreport,
      title={PaddleOCR 3.0 Technical Report}, 
      author={Cheng Cui and Ting Sun and Manhui Lin and Tingquan Gao and Yubo Zhang and Jiaxuan Liu and Xueqing Wang and Zelun Zhang and Changda Zhou and Hongen Liu and Yue Zhang and Wenyu Lv and Kui Huang and Yichao Zhang and Jing Zhang and Jun Zhang and Yi Liu and Dianhai Yu and Yanjun Ma},
      year={2025},
      eprint={2507.05595},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.05595}, 
}

@misc{cui2025paddleocrvlboostingmultilingualdocument,
      title={PaddleOCR-VL: Boosting Multilingual Document Parsing via a 0.9B Ultra-Compact Vision-Language Model}, 
      author={Cheng Cui and Ting Sun and Suyin Liang and Tingquan Gao and Zelun Zhang and Jiaxuan Liu and Xueqing Wang and Changda Zhou and Hongen Liu and Manhui Lin and Yue Zhang and Yubo Zhang and Handong Zheng and Jing Zhang and Jun Zhang and Yi Liu and Dianhai Yu and Yanjun Ma},
      year={2025},
      eprint={2510.14528},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.14528}, 
}

@misc{cui2026paddleocrvl15multitask09bvlm,
      title={PaddleOCR-VL-1.5: Towards a Multi-Task 0.9B VLM for Robust In-the-Wild Document Parsing}, 
      author={Cheng Cui and Ting Sun and Suyin Liang and Tingquan Gao and Zelun Zhang and Jiaxuan Liu and Xueqing Wang and Changda Zhou and Hongen Liu and Manhui Lin and Yue Zhang and Yubo Zhang and Yi Liu and Dianhai Yu and Yanjun Ma},
      year={2026},
      eprint={2601.21957},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.21957}, 
}

@misc{zhang2026paddleocrvl16expandingfrontierdocument,
      title={PaddleOCR-VL-1.6: Expanding the Frontier of Document Parsing with Under-Optimized Region Refinement and Progressive Post-Training}, 
      author={Zelun Zhang and Hongen Liu and Suyin Liang and Yubo Zhang and Yiqing Xiang and Jiaxuan Liu and Ting Sun and Manhui Lin and Yue Zhang and Changda Zhou and Tingquan Gao and Cheng Cui and Yi Liu and Dianhai Yu and Yanjun Ma},
      year={2026},
      eprint={2606.03264},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.03264}, 
}

English / 原文

PaddleOCR converts PDF documents and images into structured, LLM-ready data (JSON/Markdown) with industry-leading accuracy. With 70k+ Stars and trusted by top-tier projects like Dify, RAGFlow, and Cherry Studio, PaddleOCR is the bedrock for building intelligent RAG and Agentic applications.

📄 Intelligent Document Parsing (LLM-Ready)

Transforming messy visuals into structured data for the LLM era.

  • SOTA Document VLM : Featuring PaddleOCR-VL-1.6 (0.9B) , the industry's leading lightweight vision-language model for document parsing. It achieves 96.3% accuracy on OmniDocBench v1.6, leads in text, formula, and table recognition, and shows significantly enhanced capabilities in ancient documents, rare characters, seals, and charts, with structured outputs in Markdown and JSON formats.
  • Structure-Aware Conversion : Powered by PP-StructureV3 , seamlessly convert complex PDFs and images into Markdown or JSON. Unlike the PaddleOCR-VL series models, it provides more fine-grained coordinate information, including table cell coordinates, text coordinates, and more.
  • Production-Ready Efficiency : Achieve commercial-grade accuracy with an ultra-small footprint. Outperforms numerous closed-source solutions in public benchmarks while remaining resource-efficient for edge/cloud deployment.

🔍 Universal Text Recognition (Scene OCR)

The global gold standard for high-speed, multilingual text spotting.

  • 100+ Languages Supported : Native recognition for a vast global library. PP-OCRv6 supports 50 languages with a single unified model (Chinese, English, Japanese, and 46 Latin-script languages) — no model switching needed for multilingual documents.
  • Complex Element Mastery : Beyond standard text recognition, we support natural scene text spotting across a wide range of environments, including IDs, street views, books, and industrial components
  • Performance Leap : PP-OCRv6 achieves +4.6% detection and +5.1% recognition accuracy over PP-OCRv5, surpassing mainstream Vision-Language Models. 5.2× CPU inference speedup end-to-end.

🛠️ Developer-Centric Ecosystem

  • Seamless Integration : The premier choice for the AI Agent ecosystem—deeply integrated with Dify, RAGFlow, Pathway, and Cherry Studio.
  • LLM Data Flywheel : A complete pipeline to build high-quality datasets, providing a sustainable "Data Engine" for fine-tuning Large Language Models.
  • One-Click Deployment : Supports various hardware backends (NVIDIA GPU, Intel CPU, Kunlunxin XPU, and diverse AI Accelerators).

🔥 2026.06.11: Release of PaddleOCR 3.7.0

  • PP-OCRv6 highlights:

    • Accuracy boost : Medium tier achieves +4.6% detection and +5.1% recognition over PP-OCRv5_server, surpassing mainstream VLMs (Qwen3-VL-235B, GPT-5.5) with only 34.5M parameters.
    • 50 languages unified : Single model covers Chinese, English, Japanese, and 46 Latin-script languages — no model switching needed.
    • Specialized scenarios : Major improvements in digital displays, dot-matrix characters, tire prints, and industrial text recognition.
    • Faster inference : 5.2× CPU speedup (OpenVINO), 6.1× on Apple M4 (tiny), 0.13s on A100 GPU.
    • Three tiers for all scenarios : tiny (1.5M) / small (7.7M) / medium (34.5M) for edge, mobile, and server deployment.
    • Model availability : All models are available on HuggingFace and ModelScope.

2026.05.28: Release of PaddleOCR 3.6.0 2026.04.21: Release of PaddleOCR 3.5.0

  • Flexible inference backends : Seamlessly switch between Paddle static graph, Paddle dynamic graph, or Transformers. PaddleOCR is now deeply integrated with the Hugging Face ecosystem, and 20 major models support Transformers as the inference backend.
  • Office documents to Markdown : Convert common document formats such as Word, Excel, and PowerPoint into Markdown.
  • DOCX export for parsed results : The PaddleOCR-VL series, PP-StructureV3, and PP-DocTranslation now support exporting parsed results to DOCX for convenient viewing and editing in Microsoft Word.
  • Official browser inference SDK : Released PaddleOCR.js, the official browser inference SDK that supports running PP-OCRv5 directly in the browser.

2026.01.29: Release of PaddleOCR 3.4.0

  • PaddleOCR-VL-1.5 (SOTA 0.9B VLM): Our latest flagship model for document parsing is now live!
    • 94.5% Accuracy on OmniDocBench : Surpassing top-tier general large models and specialized document parsers.
    • Real-World Robustness : First to introduce the PP-DocLayoutV3 algorithm for irregular shape positioning, mastering 5 tough scenarios: Skew, Warping, Scanning, Illumination, and Screen Photography.
    • Capability Expansion : Now supports Seal Recognition , Text Spotting , and expands to 111 languages (including China’s Tibetan script and Bengali).
    • Long Document Mastery : Supports automatic cross-page table merging and hierarchical heading identification.
    • Try it now : Available on HuggingFace or our Official Website.

2025.10.16: Release of PaddleOCR 3.3.0 2025.08.21: Release of PaddleOCR 3.2.0

History Log

PaddleOCR official website provides interactive Experience Center and APIs —no setup required, just one click to experience.

👉 Visit Official Website

For local usage, please refer to the following documentation based on your needs:

🔄 Quick Overview of Execution Results

⭐ Star this repository to keep up with exciting updates and new releases, including powerful OCR and document parsing capabilities! ⭐

👩‍👩‍👧‍👦 Community

PaddlePaddle WeChat official account Join the tech discussion group

😃 Awesome Projects Leveraging PaddleOCR

PaddleOCR wouldn't be where it is today without its incredible community! 💗 A massive thank you to all our longtime partners, new collaborators, and everyone who's poured their passion into PaddleOCR — whether we've named you or not. Your support fuels our fire!

Project Name Description
Dify Production-ready platform for agentic workflow development.
RAGFlow RAG engine based on deep document understanding.
pathway Python ETL framework for stream processing, real-time analytics, LLM pipelines, and RAG.
MinerU Multi-type Document to Markdown Conversion Tool
Umi-OCR Free, Open-source, Batch Offline OCR Software.
cherry-studio A desktop client that supports for multiple LLM providers.
haystack AI orchestration framework to build customizable, production-ready LLM applications.
OmniParser OmniParser: Screen Parsing tool for Pure Vision Based GUI Agent.
QAnything Question and Answer based on Anything.
Learn more projects More projects based on PaddleOCR

This project is released under the Apache 2.0 license.

@misc{cui2025paddleocr30technicalreport,
      title={PaddleOCR 3.0 Technical Report}, 
      author={Cheng Cui and Ting Sun and Manhui Lin and Tingquan Gao and Yubo Zhang and Jiaxuan Liu and Xueqing Wang and Zelun Zhang and Changda Zhou and Hongen Liu and Yue Zhang and Wenyu Lv and Kui Huang and Yichao Zhang and Jing Zhang and Jun Zhang and Yi Liu and Dianhai Yu and Yanjun Ma},
      year={2025},
      eprint={2507.05595},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.05595}, 
}

@misc{cui2025paddleocrvlboostingmultilingualdocument,
      title={PaddleOCR-VL: Boosting Multilingual Document Parsing via a 0.9B Ultra-Compact Vision-Language Model}, 
      author={Cheng Cui and Ting Sun and Suyin Liang and Tingquan Gao and Zelun Zhang and Jiaxuan Liu and Xueqing Wang and Changda Zhou and Hongen Liu and Manhui Lin and Yue Zhang and Yubo Zhang and Handong Zheng and Jing Zhang and Jun Zhang and Yi Liu and Dianhai Yu and Yanjun Ma},
      year={2025},
      eprint={2510.14528},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.14528}, 
}

@misc{cui2026paddleocrvl15multitask09bvlm,
      title={PaddleOCR-VL-1.5: Towards a Multi-Task 0.9B VLM for Robust In-the-Wild Document Parsing}, 
      author={Cheng Cui and Ting Sun and Suyin Liang and Tingquan Gao and Zelun Zhang and Jiaxuan Liu and Xueqing Wang and Changda Zhou and Hongen Liu and Manhui Lin and Yue Zhang and Yubo Zhang and Yi Liu and Dianhai Yu and Yanjun Ma},
      year={2026},
      eprint={2601.21957},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.21957}, 
}

@misc{zhang2026paddleocrvl16expandingfrontierdocument,
      title={PaddleOCR-VL-1.6: Expanding the Frontier of Document Parsing with Under-Optimized Region Refinement and Progressive Post-Training}, 
      author={Zelun Zhang and Hongen Liu and Suyin Liang and Yubo Zhang and Yiqing Xiang and Jiaxuan Liu and Ting Sun and Manhui Lin and Yue Zhang and Changda Zhou and Tingquan Gao and Cheng Cui and Yi Liu and Dianhai Yu and Yanjun Ma},
      year={2026},
      eprint={2606.03264},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.03264}, 
}