ndl-lab/ndlocr-lite

NDLOCR-Liteアプリケーションのリポジトリ（ソースコードを含む）NDLOCR‑Lite application repository (including source code)

What it solves

NDLOCR-Lite is a lightweight OCR (Optical Character Recognition) tool designed to convert images of books and magazines into text data. It specifically addresses the need for a high-performance OCR that can run on standard home computers (like laptops) without requiring a dedicated GPU, making digital archiving more accessible on consumer hardware.

How it works

The system combines three primary modules to process images:

  1. Layout Recognition: Uses DEIMv2 (with a ViT-Tiny model distilled from DINOv3) to identify the structure of the page.
  2. Text Recognition: Employs PARSeq to recognize the actual characters within the identified areas.
  3. Reading Order Sorting: Organizes the recognized text into the correct logical sequence.

Models are trained using PyTorch and then converted to ONNX format for efficient execution. The tool can be used via a desktop application (with a screen capture mode) or a command-line interface.

Who it’s for

  • Researchers and archivists digitizing books, magazines, and historical documents.
  • Users who need to perform OCR on Japanese text (including vertical and horizontal writing) without high-end server hardware.
  • Developers looking for a customizable OCR pipeline that supports ONNX models.

Highlights

  • GPU-free operation: Optimized for fast execution on standard Windows, Mac, and Linux laptops.
  • Broad format support: Processes JPG, PNG, TIFF, JP2, BMP, and WebP files.
  • Handwriting support: Enhanced recognition for handwritten characters (as of v1.2).
  • Flexible output: Provides results in XML and JSON formats, with an optional visualization mode to highlight recognized text areas.

Related

  • Project
    arcships/light-ocr

    light‑ocr is a cross‑platform, offline OCR library for Node.js and C++. It bundles the model, PDF renderer, and a Chinese fallback font in a single npm package, provides pre‑built native binaries for macOS, Linux and Windows, and can accelerate inference with Core ML or WebGPU. The API returns text lines with confidence scores and quadrilateral coordinates, and a CLI offers quick command‑line usage. Model sizes range from a 6 MB Tiny tier to a 139 MB Medium tier, all under Apache 2.0.

  • Project
    robertknight/ocrs

    A Rust library and CLI tool for extracting text from images using neural network models to minimize the need for image preprocessing.

  • Project
    zibo-chen/ocr-rs

    A lightweight Rust OCR library based on PaddleOCR models and the MNN inference runtime, providing text detection and recognition across 50+ languages.

  • Project
    scribeocr/scribeocr

    A browser-based web application for recognizing text from images, proofreading OCR data with high precision, and creating fully digitized, ebook-style documents.

  • Project
    kha-white/mokuro

    A tool that performs text detection and OCR on Japanese manga to create selectable text overlays in a browser, facilitating the use of pop-up dictionaries for language learners.