The original article published in Japanese ( https://current.ndl.go.jp/e2891 ).
Current Awareness-E No.525
June 25, 2026
The Release of NDLOCR-Lite and Its Background
The Research and Development for Next-Generation Systems Office (the Office), Digital Information Planning Division, Digital Information Department, National Diet Library (NDL), conducts research to support the development of next-generation library systems. In recent years, the Office has placed particular emphasis on research and development related to optical character recognition (OCR) technologies, as well as on releasing the resulting software under open licenses as a means of improving access to the NDL’s digitized collections by converting them into text.
As of 2026, the Office has developed and released NDLOCR, NDL Kotenseki OCR, and their respective Lite versions. NDLOCR is designed for printed materials published from the Meiji period onward. It was developed primarily through an outsourced project during fiscal years 2021 and 2022 and has been used since January 2024 within the NDL’s internal systems, mainly for generating text from newly digitized materials. Starting in fiscal year 2022, staff of the Office internally developed and released NDL Kotenseki OCR, an OCR system for rare books and old materials containing “kuzushiji” (cursive historical Japanese script) and have continued to improve it based on feedback from researchers.
Although these OCR systems have achieved a high level of recognition accuracy, they are computationally demanding and require a graphics processing unit (GPU) for operation. While their public release was warmly welcomed, many users commented that “they are difficult to use” and that “preparing a suitable GPU environment is burdensome.”
To address these issues, the Office began in-house development of lightweight versions in fiscal year 2024. The initial target was NDL Kotenseki OCR, for which staff of the Office had an in-depth understanding of the underlying OCR technologies as a result of its in-house development. The Office proceeded with the development of a lightweight version by gradually replacing its component technologies with lighter-weight approaches, such as those used for image analysis in automotive cameras. The completed lightweight version was named “NDL Kotenseki OCR-Lite” and released in November 2024. It received a strong response from humanities researchers as well as from NDL staff.
The key feature of NDLOCR-Lite is its lightweight design, which enables it to run quickly on a standard laptop while maintaining character recognition performance comparable to that of NDLOCR. It can operate with approximately 1 GB of available memory, and all processing is completed locally without any communication with external systems. A desktop application for Windows, macOS and other operating systems is provided, allowing users to select images, execute OCR, and review results using only mouse operations. Compared with NDL Kotenseki OCR-Lite, NDLOCR-Lite generally handles documents containing a much larger number of characters, which increases the amount of data to be processed and can result in longer processing times. To address this issue, a mechanism was developed that automatically identifies areas requiring computationally intensive processing and areas where lightweight processing is sufficient, and efficiently selects and applies processing methods according to the target documents. This enables high-speed OCR processing even on computers with relatively modest specifications.
Installing the application on a user’s computer is simple: users only need to download the ZIP file corresponding to their operating system from the GitHub Releases page, extract the files, and launch the executable. Care should be taken not to include Japanese or other full-width characters in the name of the folder to which the files are extracted. OCR output can be exported in a variety of formats, including TXT, JSON, XML, TEI, and PDF with a hidden text layer. Users can also copy and paste OCR results displayed in the user interface (UI).
To meet operational requirements within the NDL, NDLOCR-Lite also includes a Crop & OCR mode, which performs OCR only on a user-selected area of an image, and a Capture mode, which allows users to select an area of the displayed screen and perform OCR directly on it.
NDLOCR-Lite can also be invoked from the command line for batch processing. Furthermore, because its OCR models are distributed in the highly portable ONNX (Open Neural Network Exchange) format, it can be easily integrated into other programs to develop custom applications. One example is NDLOCR-Lite Web, a browser-based derivative application of NDLOCR-Lite developed by Dr. Hashimoto Yuta (National Museum of Japanese History / adjunct researcher at the NDL).
In April 2026, the character recognition performance of NDLOCR-Lite was further enhanced using a training dataset developed through an outsourced project. The dataset was created based on handwritten constitutional and political documents and typewritten Occupation-period materials selected in fiscal year 2025 in cooperation with the NDL’s collection management divisions. The Office will continue to improve recognition performance and add new features in collaboration with relevant divisions within the NDL and relevant external organizations.
In addition, as part of ongoing efforts to explore further tools to improve workflow efficiency, the Office is also conducting proof-of-concept experiments that integrate NDLOCR-Lite with a locally hosted large language model (LLM) running on the NDL’s internal servers. This enables operations such as document structuring, information extraction, machine translation, and summarization to be executed locally through NDLOCR-Lite’s integrated workflow without any communication with external services.
By making the technologies developed by the Office, including NDLOCR-Lite, available for use as widely and freely as possible, we hope to contribute to the advancement of Japanese-language OCR technology as a whole and to improve access to Japanese-language materials in digital archives. We encourage everyone to try NDLOCR-Lite.
Aoike Toru
Research and Development for Next-Generation Systems Office,
Digital Information Planning Division,
Digital Information Department,
National Diet Library
Translated by
Research and Information Section,
Library Support Division,
Kansai-kan of the National Diet Library
*References are not translated.
Ref:
“NDLOCR-Liteの公開について”. NDL Lab. 2026-02-24.
https://lab.ndl.go.jp/news/2025/2026-02-24/
ndl-lab. “ndlocr-lite”. GitHub.
https://github.com/ndl-lab/ndlocr-lite/
ndl-lab. “ndlocr-lite, Releases”. GitHub.
https://github.com/ndl-lab/ndlocr-lite/releases
“NDLOCR-Liteの使い方”. NDL Lab.
https://lab.ndl.go.jp/data_set/ndlocrlite-usage/
ndl-lab. “ndlocr_cli”. GitHub.
https://github.com/ndl-lab/ndlocr_cli
ndl-lab. “ndlkotenocr_cli”. GitHub.
https://github.com/ndl-lab/ndlkotenocr_cli
ndl-lab. “ndlkotenocr-lite”. GitHub.
https://github.com/ndl-lab/ndlkotenocr-lite
“Guidelines for Electronic Text Encoding and Interchange”. TEI.
https://tei-c.org/release/doc/tei-p5-doc/en/html/index.html
NDLOCR-Lite Web.
https://ndlocr-liteweb.netlify.app/
