ndl-lab/ndlocr-lite

What it solves

NDLOCR-Lite provides a lightweight Optical Character Recognition (OCR) solution for digitizing books and magazines. It is specifically designed to run on standard home computers (laptops) and common operating systems without requiring a GPU, enabling fast text extraction from digital images of archival materials.

How it works

The system combines three primary modules to process images into text:

1. Layout Recognition: Uses DEIMv2 to identify the structure of the document.
2. String Recognition: Uses PARSeq to recognize the actual characters.
3. Reading Order Sorting: Organizes the recognized text into the correct sequence.

Models are trained using PyTorch and then converted to ONNX format for efficient execution. The application can be used via a desktop GUI (supporting Windows, Mac, and Linux) or a command-line interface (CLI) using Python.

Who it’s for

It is intended for users who need to digitize printed or handwritten Japanese text from images but lack high-end hardware (like dedicated GPUs) or those who prefer a simple desktop application for archival work.

Highlights

• Hardware Efficient: Designed to run fast on general-purpose CPUs without needing a GPU.
• Multi-Platform: Compatible with Windows 11, macOS (including Apple M4), and Ubuntu 22.04.
• Versatile Input: Supports various image formats including JPG, PNG, TIFF, JP2, BMP, and WebP.
• Handwriting Support: Enhanced recognition for handwritten characters (as of v1.2).
• Flexible Output: Can output results in XML or JSON formats.