PaddleOCR is a comprehensive optical character recognition toolkit and document AI engine developed by PaddlePaddle that converts PDF documents and images into structured, machine-readable data suitable for large language models.
The repository's core functionality centers on two primary capabilities. First, it provides intelligent document parsing through its PaddleOCR-VL series models, with the latest version PaddleOCR-VL-1.6 achieving 96.3% accuracy on OmniDocBench v1.6. This lightweight 0.9-billion-parameter vision-language model excels at recognizing text, formulas, and tables while handling challenging scenarios such as ancient documents, rare characters, seals, and charts. The toolkit outputs structured data in both Markdown and JSON formats. Second, it offers universal text recognition across 100+ languages through PP-OCRv6, which supports 50 languages with a single unified model, eliminating the need for model switching when processing multilingual documents. The latest version achieves 4.6% improvement in detection accuracy and 5.1% improvement in recognition accuracy compared to PP-OCRv5, while delivering 5.2 times faster CPU inference speeds.
The PP-StructureV3 algorithm provides structure-aware conversion capabilities, transforming complex PDFs and images into Markdown or JSON with fine-grained coordinate information including table cell positions and text locations. The toolkit also includes PP-DocLayoutV3 for handling irregular document shapes across five challenging scenarios: skew, warping, scanning artifacts, illumination variations, and screen photography.
The repository supports flexible deployment across multiple hardware backends including NVIDIA GPUs, Intel CPUs, Kunlunxin XPUs, and various AI accelerators. Recent releases have expanded functionality to include office document conversion to Markdown, DOCX export for parsed results, and browser-based inference through the PaddleOCR.js SDK. The toolkit provides three model tiers—tiny at 1.5 million parameters, small at 7.7 million parameters, and medium at 34.5 million parameters—enabling deployment across edge devices, mobile platforms, and server environments. Models are distributed through HuggingFace and ModelScope repositories, facilitating integration with the broader machine learning ecosystem. The project maintains documentation in multiple languages including English, Simplified Chinese, Traditional Chinese, Japanese, Korean, French, Russian, Spanish, and Arabic, reflecting its global user base.
