Android: Google Play
iOS: App Store
At the noon of Japan Standard Time on August 30, 2021, the Android and iOS version of the app was officially released. The app is free of charge. Install the app from Google Play or App Store, and use it on your smartphones or tablets. [Read more in Japanese]
Please use the hashtag "#miwoapp" to share information about the miwo app.
Before using the app
The AI kuzushiji recognition is not perfect. The result of kuzushiji recognition may contain errors, and users should check it by themselves. CODH is not accepting inquiry about the error of kuzushiji recognition. If you would like to obtain the result of kuzushiji transcription without errors, please ask for transcription services provided by companies and individuals.
The AI kuzushiji recognition sends an image to the server operated by CODH, but CODH will not save the image sent to the server nor the result of recognition.
- You may be able to solve the problem by updating your OS to the latest version.
- A device that was sold more than four to five years ago is more likely to cause problems both on iOS and Android. We are working to fix these issues.
- The app needs permission to access to the camera and the album. When you happened to reject the permissions and could not use the app, please remove the app and install it again.
If the actions above cannot solve the problem, please tell us your situation from the page below. Information about the type of device (model) of the smartphone or the table, or the version of the OS is much appreciated.
The prototype of the miwo app was released for the exhibition program Letter shape @ Keio Museum Commons (KeMCo) from April 19 to June 18, 2021.
Kuzushiji and AI
It is said that only a few thousand people (about 0.01% of the Japanese population) can read Kuzushiji fluently. In order to make historical materials written in Kuzushiji more accessible to the general public, we need to transcribe the abundant amount of documents which would take very long time because of limited human resource. The question is: is it possible to use AI (artificial intelligence) to help transcribe the materials?
The ROIS-DS Center for Open Data in the Humanities (CODH) has developed “KuroNet” Kuzushiji recognition system which can automatically find characters and transcribe text from image using AI object detection technology. KuroNet is trained on the “Kuzushiji dataset” created by the National Institute of Japanese Literature and is skilled at recognizing Kuzushiji characters especially in Edo period woodblock printed books.
CODH is also developing the smartphone app "miwo" so that anyone can use KuroNet easily. The name “miwo” comes from the 14th chapter of The Tale of Genji “miwotsukushi”, referring to waterway signs. Just as the miwotsukushi is a guide for boats in the sea, we aim to make our "miwo" app as a guide for traveling the ocean of historical documents.
AI Kuzushiji Recognition
The miwo app uses two kuzushiji recognition models, KuroNet developed by CODH and the model developed by tascj who won the Kaggle Competition of Kuzushiji Recognition. Those AI models were trained on Kuzushiji Dataset, created by National Institute of Japanese Literature and curated by CODH.
- Tarin Clanuwat
- Asanobu Kitamoto
- Mikel Bober-Irizar
- Alex Lamb
- Siyu Han
miwo app is free of charge to support humanities researches and promote classical Japanese culture. The servers (including GPUs) for the miwo app is operated by ROIS-DS Center for Open Data in the Humanities. The development of the app and the research of AI kuzushiji recognition is also supported by the following grants.
- End-to-end Pre-modern Japanese Kuzushiji Recognition with Deep Learning, JSPS kakenhi Grant-in-Aid for Early-Career Scientists (19K13085), PI Tarin Clanuwat
- 資料調査のためのオンデバイスくずし字認識, JST ACT-X AI活用で挑む学問の革新と創成(JPMJAX20A4), PI Tarin Clanuwat
- Data-Driven Reconstruction and Integrated Analysis of the Past World Using the Infrastructure for Historical Big Data, JSPS Kakenhi Grant-in-Aid for Scientific Research (A) (19H01141), PI Asanobu Kitamoto
Since the release, 129,664 images have been recognized as of 2021-10-23T16:24:19+09:00.