DOCSEARCHbyAlgolia
Submit Docs

PaddleOCR: Multilingual OCR and Document AI Toolkit

repository·main·Indexed 11 days ago

https://github.com/paddlepaddle/paddleocr

A comprehensive OCR and document parsing toolkit based on PaddlePaddle that converts images and PDFs into structured formats like JSON and Markdown. It provides SDKs for Python, TypeScript, and Go, and supports deployment on Android and bare metal Arm Cortex-M55. The toolkit includes capabilities for text detection, classification, and recognition, as well as tools for training and evaluating DBNet models.

Tokens
482.2K
Snippets
1.1K
Records
1.8K
Agent score
95%

What's inside PaddleOCR

  1. Overview of PaddleOCR

    main

    PaddleOCR is a leading OCR toolkit and document AI engine designed to convert documents and images into structured, LLM-ready data (JSON/Markdown) with industry-leading accuracy. It is widely used in RAG (Retrieval-Augmented Generation) and AI agent applications, integrated into projects like Dify, RAGFlow, and Cherry Studio.

    Key Capabilities

    1. Intelligent Document Parsing (LLM-Ready)

    Converts unstructured visual data into structured formats like Markdown and JSON.

    • PaddleOCR-VL-1.6 (0.9B): A lightweight Vision-Language Model (VLM) optimized for text, formulas, and tables. It excels at handling ancient documents, rare characters, seals, and charts.
    • PP-StructureV3: Enables seamless conversion of complex PDFs and images into Markdown or JSON. Unlike the VL series, it provides detailed coordinate information for table cells and text.
    • Production Efficiency: Designed for high accuracy with a small footprint, suitable for edge or cloud deployment.

    2. Universal Text Recognition (Scene OCR)

    Standard for high-speed, multilingual text detection.

    • Multilingual Support: Native recognition for over 100 languages. The PP-OCRv6 model provides unified support for 50 languages (including Chinese, English, Japanese, and 46 Latin-based languages) within a single model.
    • Natural Scene Text Detection: Supports diverse environments such as ID cards, street scenes, books, and industrial parts.
    • Performance: PP-OCRv6 offers significant improvements over PP-OCRv5, including a +4.6% increase in detection accuracy and +5.1% in recognition accuracy, with 5.2x faster CPU inference.

    3. Developer Ecosystem

    • Seamless Integration: Deeply integrated with AI agent ecosystems like Dify, RAGFlow, Pathway, and Cherry Studio.
    • LLM Data Flywheel: Provides a pipeline for building high-quality datasets for LLM fine-tuning.
    • One-Click Deployment: Supports various hardware backends including NVIDIA GPU, Intel CPU, Kunlunxin XPU, and other AI accelerators.

  2. Overview of PP-OCRv6 features

    main

    Released in June 2026, PP-OCRv6 offers significant improvements over previous versions:

    • Accuracy: The medium tier achieves +4.6% detection and +5.1% recognition accuracy compared to PP-OCRv5_server, outperforming mainstream VLMs (like Qwen3-VL-235B or GPT-5.5) with only 34.5M parameters.
    • Language Support: Integrated support for 50 languages (including Chinese, English, Japanese, and 46 Latin languages) in a single model without needing model switching.
    • Specialized Scenarios: Enhanced recognition for digital displays, dot matrix characters, tire printing, and industrial characters.
    • Inference Speed: Accelerated performance via OpenVINO (5.2x on CPU), Apple M4 (6.1x for tiny), and A100 GPU (0.13s).
    • Model Tiers: Three models available for different hardware: tiny (1.5M), small (7.7M), and medium (34.5M).
    • Model Availability: Models can be downloaded from HuggingFace or ModelScope.

  3. Overview of PaddleOCR features and solutions

    main

    PaddleOCR provides a complete pipeline for OCR tasks, including data production, model training, compression, inference, and deployment. Key components include:

    • PP-OCR: The core series of industrial-grade OCR models.
    • PP-Structure: A solution for document analysis, including layout analysis, table recognition, Semantic Entity Recognition (SER), and Relation Extraction (RE).
    • PP-ChatOCR: An advanced OCR solution (refer to the external link for details).
    • PaddleX: A high-efficiency development platform that offers two modes:
      • Toolbox mode: Allows tuning key hyperparameters without writing code.
      • Developer mode: Supports single-model training, deployment, and multi-model serial inference with low code, compatible with both cloud and local terminals.

  4. Overview of DB and DB++ text detection algorithms

    main

    DB (Differentiable Binarization) and DB++ (Differentiable Binarization with Adaptive Scale Fusion) are real-time scene text detection algorithms.

    Key Models and Backbones:

    • DB with ResNet50_vd: High precision (86.41% on ICDAR2015).
    • DB with MobileNetV3: Lightweight option (77.29% on ICDAR2015).
    • DB++ with ResNet50: Improved performance (90.89% on ICDAR2015; 92.92% on TD_TR).

    Note: Models trained on ICDAR2015 (primarily English) may perform poorly on Chinese text images.

  5. Overview of the SEED algorithm

    main

    SEED (Semantics Enhanced Encoder-Decoder Framework for Scene Text Recognition) is a scene text recognition algorithm based on the CVPR 2020 paper. It uses a semantics-enhanced approach to improve recognition accuracy.

    Model Details:

    • Backbone: Aster_Resnet
    • Average Accuracy: 85.20%
    • Configuration File: configs/rec/rec_resnet_stn_bilstm_att.yml
    • Training Datasets: MJSynth and SynthText
    • Evaluation Datasets: IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE

  6. Overview of PaddleOCR 3.x Major Features

    main

    PaddleOCR 3.x provides a suite of advanced OCR and document intelligence capabilities:

    • PaddleOCR-VL: A resource-efficient Vision-Language Model (0.9B parameters) designed for multilingual document parsing. It supports 109 languages and excels at recognizing complex elements like text, tables, formulas, and charts.
    • PP-OCRv6: A universal multi-language text recognition system. A single model supports 50 languages (including Chinese, English, Japanese, and 46 Latin-script languages). It offers three tiers (tiny, small, and medium) ranging from 1.5M to 34.5M parameters for various deployment scenarios (edge to server).
    • PP-StructureV3: Specialized for complex document parsing. It converts PDFs and document images into Markdown and JSON files while preserving the original hierarchical structure and layout.
    • PP-ChatOCRv4: An intelligent information extraction tool that integrates ERNIE 4.5, allowing users to ask questions about documents and receive accurate answers based on the content.

  7. Overview of PaddleOCR Official API SDKs

    main

    The PaddleOCR Official API SDKs are client-side libraries used to call the hosted PaddleOCR official API services.

    Important Note: These SDKs do not perform OCR inference locally and do not load local models. They act as interfaces to communicate with the hosted service.

    Available SDKs include:

    • Python SDK: Integrated into the main paddleocr package.
    • TypeScript SDK: Available in the typescript directory.
    • Go SDK: Available in the go directory.

  8. Overview of Rosetta Text Recognition

    main

    Rosetta is a large-scale system for text recognition in images, based on the paper Rosetta: Large Scale System for Text Detection and Recognition in Images. It is trained on MJSynth and SynthText datasets and can be evaluated on datasets like IIIT, SVT, IC03, IC13, IC15, and SVTP.

    Available models include:

    • Resnet34_vd: Avg Accuracy ~79.11% (Config: configs/rec/rec_r34_vd_none_none_ctc.yml)
    • MobileNetV3: Avg Accuracy ~75.80% (Config: configs/rec/rec_mv3_none_none_ctc.yml)

  9. Overview of the Document Image Preprocessing Pipeline

    main

    The Document Image Preprocessing Pipeline is designed to improve document image quality for OCR and digital document management. It integrates two primary optional modules:

    1. Document Image Orientation Classification: Automatically identifies the document's orientation (0°, 90°, 180°, or 270°) to ensure correct alignment for downstream tasks.
    2. Text Image Unwarping: Corrects geometric distortions caused by camera angles or scanning processes, restoring the document to its original shape and proportions.

    The pipeline supports flexible deployment across various hardware and programming languages, and allows for secondary development by training and integrating custom models.

  10. Overview of PP-OCRv6 Model Family

    main

    PP-OCRv6 is a universal text recognition solution featuring a three-tier model family designed for different deployment scenarios. It uses the PPLCNetV4 unified backbone and supports up to 50 languages (including Simplified Chinese, Traditional Chinese, English, Japanese, and 46 Latin-script languages) within a single unified model.

    Model Tiers:

    • Tiny: 1.5M parameters. Targets edge/IoT devices. Supports 49 languages (excludes Japanese).
    • Small: Targets mobile/desktop scenarios.
    • Medium: 34.5M parameters. Targets server scenarios and large-scale data pipelines. Supports 50 languages.

    PP-OCRv6 is designed to be more accurate and faster than previous versions (PP-OCRv5) and outperforms many large Vision-Language Models (VLMs) in specialized OCR tasks while maintaining a much smaller parameter footprint.

  11. Overview of PaddleOCR features

    main

    PaddleOCR is a leading OCR toolkit and document AI engine designed to convert PDF documents and images into structured, LLM-friendly formats like JSON and Markdown. It is widely used in RAG (Retrieval-Augmented Generation) and Agentic applications.

    Key Capabilities

    1. Intelligent Document Parsing (LLM-Ready)

    Converts visual document information into structured data for Large Language Models.

    • PaddleOCR-VL-1.6 (0.9B): A state-of-the-art lightweight Document Vision-Language Model (VLM). It excels at recognizing text, formulas, and tables, with enhanced capabilities for ancient books, rare characters, seals, and charts. It supports Markdown and JSON output.
    • PP-StructureV3: Drives layout structure analysis, converting complex PDFs and images into Markdown or JSON. Unlike the VL series, it provides fine-grained coordinate information (e.g., table cell coordinates, text coordinates).
    • Production Efficiency: High accuracy with small model sizes, optimized for edge computing and cloud deployment.

    2. Universal Text Recognition (Scene OCR)

    Fast and precise multi-language text detection and recognition.

    • 100+ Languages Supported: The PP-OCRv6 single model supports 50 languages (including Chinese, English, Japanese, and 46 Latin-based languages) in a single model without switching.
    • Complex Scenarios: Supports natural scene text detection for IDs, street scenes, books, and industrial parts.
    • Performance: PP-OCRv6 offers a 4.6% improvement in detection accuracy and a 5.1% improvement in recognition accuracy over PP-OCRv5, with 5.2× CPU inference acceleration.

    3. Developer Ecosystem

    • Seamless Integration: Deeply integrated with Dify, RAGFlow, Pathway, and Cherry Studio.
    • LLM Data Pipeline: Provides a data engine for building high-quality datasets to fine-tune LLMs.
    • One-click Deployment: Supports multiple hardware backends including NVIDIA GPU, Intel CPU, Kunlun XPU, and various AI accelerators.

  12. Overview of PP-Structure intelligent document analysis

    main

    PP-Structure is an intelligent document analysis system designed for document understanding tasks like layout analysis and table recognition. The system operates via two primary pipelines:

    1. Layout Analysis Pipeline:

      • Image Direction Correction: Identifies and corrects the orientation of the document image.
      • Layout Analysis: Divides the image into functional areas (e.g., text, table, figure, title, formula).
      • Area-Specific Processing: Table areas are sent to the form recognition module for structured recognition, while text areas are sent to the OCR engine.
      • Layout Recovery: Restores the analyzed content into a Word or PDF file that maintains the original layout.

    2. Key Information Extraction (KIE) Pipeline:

      • OCR: Extracts raw text content.
      • Semantic Entity Recognition (SER): Identifies semantic entities within the image.
      • Relationship Extraction (RE): Determines the correspondence/relationships between identified semantic entities.

    PP-StructureV2 modules can be used independently (e.g., using only table recognition) or collocated to form a complete pipeline.