Recognition of Japanese Historical Hand-Written Characters Based on Object Detection Methods

We consider the recognition problem of Japanese historical handwritten characters called "Kuzushiji". Unlike modern characters, Kuzushiji characters are harder to recognize partly because many of them are connected and not separated by spaces without any segmentation information. We propose two methods for segmentation and recognition of Kuzushiji characters. The first method learns segmentation rules and character classifiers simultaneously from data sets with character labels and segmentation information. Second method is for segmentation and can be used with any single character recognizer. Our methods outperform other baselines and achieve the state-of-the-art accuracy on both segmentation and recognition tasks on data sets of three consecutive Kuzushiji characters.