OCR Open Source Project In-Depth Inventory: Top 10 Not to Miss in 2025

OCR technology is capable of converting textual information in an image into editable and processable text data. In simple terms, it recognizes and extracts text from images.

Next, we'll review the 10 most starred OCR open source projects on GitHub to provide you with a thorough guide to choosing an OCR tool.

01 GOT-OCR 2.0: End-to-End Multimodal OCR Models

GOT-OCR 2.0 is an open source end-to-end multimodal OCR model with a model size of only 1.43 GB. It not only recognizes and extracts text, but also processesMath formulas, molecular formulas, diagrams, sheet music, geometric shapesThis has greatly broadened the scope of application of OCR technology.

Model Features:

• Multimodal support: In addition to regular text, it can handle a wide range of complex content.
• Lightweight models: With a model size of only 1.43 GB, it is easy to deploy.
• End-to-end recognition: No need for complex pre-processing and post-processing processes.

Advantage: GOT-OCR 2.0 has obvious advantages in handling complex scenarios and diversified contents, and is suitable for application scenarios that need to handle multiple types of documents.

It's currently got 7.2K Stars on GitHub!

Open source address: https://github.com/Ucas-HaoranWei/GOT-OCR2.0

02 InternVL: a powerful open source multimodal model

InternVL is an open-source multimodal macromodel developed by the OpenGVLab team that aims to provide a close approximation of the GPT-4V and Gemini An alternative to the performance of commercial models such as Pro.

Although InternVL belongs to the visual big model, the application scene is more extensive, such as image understanding, not the vertical model of the OCR field, but it can be backward compatible with the OCR extraction of the text of the scene. There are many excellent open source visual models, this article does not list them all, only to InternVL as an example.

Model Features:

• Multimodal capabilities: Supports a wide range of tasks such as image understanding and visual quizzing.
• High performance: Approaching the performance of a commercial model.
• Open Source Open: Convenient for developers for secondary development and customization.

Advantage: InternVL, as a visual macromodel, has advantages in processing complex images and understanding image content, and also meets the basic needs of OCR.

It has received 7.2K Stars so far.

Open source address: https://github.com/OpenGVLab/InternVL

03 olmOCR: PDF document structured processing experts

olmOCR is developed by AllenAI and is focused on PDF document linearizationA toolkit that converts complexly laid out PDFs into structured text suitable for Large Language Model (LLM) training.

Its core objective is to generate coherent text data by efficiently handling PDF issues such as mixed text and graphics, multi-column layout, etc., to enhance LLM's ability to understand documents in real-world scenarios.

Technical Details:

• Layout Analysis: Accurately recognize multi-column layouts of text, images, tables, etc. in PDF.
• Text linearization: Convert complex layouts into linear text sequences suitable for LLM processing.
• Content reorganization: Solve problems such as cross-page, cross-column, etc., to ensure the coherence of the text.

Application Scenarios:

• Analysis of academic papers: Quickly extract key information from your paper.
• Legal document processing: Structured extraction of document content such as contracts, judgments, etc.
• Financial Statement Analysis: Automated extraction of financial data and key metrics.

Required configuration is an up-to-date NVIDIA GPU (tested on RTX 4090, L40S, A100, H100) with at least 20 GB of GPU RAM and 30 GB of available disk space.

It has received 9.8K Stars so far!

Open source address: https://github.com/allenai/olmocr
Online demo: https://olmocr.allenai.org/

04 Zerox: AI-Driven Structured Document Conversion Tool

Zerox It is an AI-driven document extraction tool developed by the Omni-AI team that converts documents in PDF, image, Docx, etc. into structured Markdown files.

Advantage:

• No training required: Unlike traditional OCR tools, Zerox can handle complex layouts without having to train the model in advance.
• Generate structured content directly: Implement OCR based on visual models (e.g., GPT-4o-mini) and generate structured content directly.
• Retain the logical structure: Recognize the columnar layout of academic papers, code blocks in technical documents, contract forms, test paper formulas, etc., and generate neat Markdown.
• Compare to traditional OCR : Zerox omits the traditional steps of layout analysis, table structure reduction, etc. and outputs Markdown results directly.

Currently getting 10.3K Star!

Open source address: https://github.com/getomni-ai/zerox
Experience address: https://getomni.ai/ocr-demo

05 Surya: Recognizing Multilingual Text and Complex Document Structures

Surya Specialized in recognizing multilingual text and complex document structures, especially good at table recognition.

Keywords: line-level text detection, layout analysis (detection of tables, images, captions, etc.), reading order detection, table recognition (detection of rows/columns), LaTeX OCR

Key Features:

1. Multi-language support: Support for more than 90 languages, including complex scripts such as Chinese, Japanese, and Arabic, as well as mainstream languages such as English and Spanish, making it suitable for document processing in globalized scenarios.
2. Form recognition optimization: Can accurately recognize the rows, columns and cell structure of the table , including rotating or complex layout of the table , performance better than the current mainstream open source models (such as Table Transformer).
3. Complex document parsing: It can detect the title, pictures, paragraphs and other elements in the document, and intelligently determine the reading order to avoid the output content confusion.

Application Scenario Example:

• Digitization of multilingual documents: Multilingual contracts, reports, etc. are handled by multinational companies.
• Digitization of historical archives: Handle historical documents containing complex tables and layouts.
• Scientific data extraction: Extracting tabular data from academic papers.

Surya supports CPU/GPU operation and significantly improves recognition speed through batch processing and image pre-processing optimizations (e.g., denoising, grayscaling) for enterprise-level document digitization needs.

It's currently got 16.8K Stars on GitHub!

Open source address: https://github.com/VikParuchuri/surya

06 OCRmyPDF: Add a searchable text layer to scanned PDFs

This open source tool , designed for scanned PDF files ( i.e. PDF is all images , images in the text can not be copied ) to add a searchable , copyable text layer .

Application Scenarios:

• Digitization of archives: Convert scanned paper documents to searchable PDF.
• Accessibility: Accessible PDF documents for the visually impaired.
• Information Retrieval: Easy to find information from a large number of scanned documents.

Advantage:

• Precise identification: Supports more than 100 languages using the Tesseract OCR engine.
• Image Optimization: Automatically corrects skewed pages and rotated error pages to improve recognition rates.
• Batch processing: Efficiently process thousands of pages of documents with multi-core CPU acceleration.

OCRmyPDF offers a significant advantage in processing scanned PDFs, is easy to install and use, and is compatible with Linux, Windows, macOS, and Docker. oCRmyPDF provides a more convenient solution than other tools that require manual processing of scanned documents.

Currently it has 20.7K Stars on GitHub!

When opening image-based PDF, you will find that the text on the image cannot be copied and searched.OCRmyPDF can embed OCR text layer under the image, supporting high-precision copying and searching.

Open source address: https://github.com/ocrmypdf/OCRmyPDF
Access Documentation: https://ocrmypdf.readthedocs.io/en/latest/

07 Marker: PDF, images and other multi-format document conversion

Marker is an efficient document conversion tool developed by Vik Paruchuri, which can quickly convert PDF, images, Office documents and EPUB formats to Markdown, JSON or HTML.

Advantage: Marker It excels in parsing complex content (e.g. tables, math formulas, code blocks) with high accuracy and excellent processing speed, supports GPU acceleration, and outperforms similar cloud services (e.g. Llamaparse, Mathpix).

Applications:

• Academic paper conversion: Convert PDF papers to Markdown for easy editing and citation.
• Technical Documentation Generation: Convert documents containing code and diagrams into an easy-to-publish HTML format.
• Data Extraction: Extract table and form data into JSON format for easy subsequent processing.

Marker can invoke large language models (e.g. Gemini, Ollama) to optimize results, such as cross-page table merging, formula formatting, and form data extraction.

It's currently got 22.8K Stars on GitHub.

Open source address: https://github.com/vikParuchuri/marker

08 EasyOCR: Multilingual Text Recognition Tool Library

EasyOCR It is an open source OCR tool library developed by JaidedAI, which inputs an image and returns the extracted text, the coordinates of the corresponding location, and the confidence level.

Features:

• Multi-language support: Support for more than 80 languages and multiple writing systems (e.g., Chinese, Latin, Arabic).
• Ready-to-use: Provides pre-trained models for rapid deployment without additional training.
• Flexible input: Supports multiple input forms such as images, byte streams, URLs, and more.
• Simplicity API: Output text content, position, and confidence via a concise API.
• CPU/GPU compatible: The operating environment can be flexibly selected according to the hardware conditions.

Model Training: EasyOCR is based on the PyTorch deep learning framework and uses a CRNN (Convolutional Recurrent Neural Network) model structure combined with a CTC (Connectionist Temporal Classification) loss function for training.

Application Scenarios:

• Multilingual document recognition: Ideal for working with documents containing multiple languages.
• Natural scene text recognition: It can be used to recognize text in natural scenes such as street signs and license plates.
• Mobile OCR: The model is lightweight and suitable for deployment on mobile.

EasyOCR balances developer-friendliness and industrial-grade application requirements for OCR scenarios such as multilingual documents and natural scene text.

It currently has 26K Stars on GitHub.

Open source address: https://github.com/JaidedAI/EasyOCR
Demo address: https://www.jaided.ai/documentai/demo

09 Umi-OCR: Install-and-use offline OCR software

This is a free, open source, offline OCR text recognition software, supports Windows 7+ x64 and Linux x64 systems, no need to network, download and run locally.

Keywords: local software unzip and run offline; screenshot OCR; batch OCR;

Advantage:

• Running offline: No internet connection is required to protect user privacy.
• Easy to use: Provides a graphical interface for easy operation.
• Feature-rich: Support screenshot OCR, batch OCR and many other functions.
• Compare this to other offline tools: Features easy installation and no need to configure the operating environment.

So far it has earned 30.8K Stars.

Open source address: https://github.com/hiroi-sora/Umi-OCR

10 Tesseract: Ancient Gods of the OCR Realm

Tesseract is a powerful and widely used open source OCR engine that converts text in images into editable text.

Historical Context:

• Developed by Hewlett-Packard Laboratories between 1985 and 1994.
• It was ported to Windows after 1996.
• HP open-sourced it in 2005.
• Sponsored by Google, it is one of the more recognizable open source OCR systems.

Technical characteristics:

• Deep learning techniques: Character recognition using advanced deep learning techniques (e.g., convolutional neural networks) is highly accurate and performs well especially when dealing with better quality scanned images.
• Multi-language support: Supports text recognition in over 100 languages.

Compare it to other engines: Tesseract has a long history, an active community, and is well documented, but may not be as good as some of the emerging OCR engines at handling complex layouts and low-quality images.

There is also a JavaScript version of Tesseract OCR: Tesseract.js, but after actual testing, it was found that the JS version doesn't support Chinese very well.

It has received 65.3K Stars on GitHub so far.

Open source address: https://github.com/tesseract-ocr/tesseract
Open source address: https://github.com/naptha/tesseract.js