“miwo” AI Kuzushiji Recognition Application for Document Examination

Tarin Clanuwat
Asanobu Kitamoto
Proceeding of IPSJ Humanities and Computer Symposium 2021
Google Scholar

Abstract

Kuzushiji character recognition has made rapid progress in recent years. The recognition models such as KuroNet and the Kaggle competition winning models have achieved over 90% accuracy on the Kuzushiji dataset. In order to make these Kuzushiji recognition models available for anyone, we developed an application called “miwo” that can take a picture of the Kuzushiji document get the recognition result easily from Kuzushiji recognition system on CODH server. We can say that miwo app has taken Kuzushiji recognition models into the production phase. In this paper, we explain about the development and release of the miwo app. Then, we discuss the Kuzushiji recognition research from the perspective of machine learning project life cycle and machine learning operations. Finally, we propose ways to improve Kuzushiji recognition system in the future

Research Areas