Recognition of Japanese handwritten characters with Machine learning techniques

Abstract

The recognition of Japanese handwritten characters has always been a challenge for researchers. A large number of classes, their graphic complexity, and the existence of three different writing systems make this problem particularly difficult compared to Western writing. For decades, attempts have been made to address the problem using traditional OCR (Optical Character Recognition) techniques, with mixed results. With the recent popularization of machine learning techniques through neural networks, this research has been revitalized, bringing new approaches to the problem. These new results achieve performance levels comparable to human recognition. Furthermore, these new techniques have allowed collaboration with very different disciplines, such as the Humanities or East Asian studies, achieving advances in them that would not have been possible without this interdisciplinary work. In this thesis, these techniques are explored until reaching a sufficient level of understanding that allows us to carry out our own experiments, training neural network models with public datasets of Japanese characters. However, the scarcity of public datasets makes the task of researchers remarkably difficult. Our proposal to minimize this problem is the development of a web application that allows researchers to easily collect samples of Japanese characters through the collaboration of any user. Once the application is fully operational, the examples collected until that point will be used to create a new dataset in a specific format. Finally, we can use the new data to carry out comparative experiments with the previous neural network models.