Collection of Online and Offline Handwritten Japanese Charcters and Handwriting Classification Using the Data

This paper describes the establishment of an online and offline Japanese handwritten character dataset and experiments on handwriting identification. Handwriting data were collected for forensic writer identification with 5-time repetition to examine intra-individual constancy, which was also as important as inter-individual difference. Handwritten samples were collected from four hundred and one participants. Online data included the data on the pen-tip location, the pen pressure and the time at each measuring point. The dataset had 702 characters which included uppercase and lowercase Latin alphabet letters, numerals, Hiragana, Katakana and Kanji characters. Three kinds of writer classification experiments were done using the offline data in the dataset. Image data used were 26 uppercase and 26 lowercase Latin alphabet letters, 26 Hiragana characters and 26 Kanji characters written by 10 writers. In each experiment, four samples out of 5 samples per writer and per letter or character were used for training and rest of one sample was used for test. LeNet was applied to the classification. Correct classification ratio was about 80% except lowercase letters. Lowercase letters showed 70%. Correct ratio improved by the increase in the number of the training data. Some participants showed high correct ratio and their samples were unique regarding the size, the position or the line quality. These results suggested the importance of the sample size in the classification using a Convolutional Neural Network rather than the equality of writing conditions and that the existence of the characteristics consistent to all handwriting of a writer.

I. Introduction

There have been several kinds of Japanese handwriting datasets such as the ETL Character Databases [1] and Kuchibue and Nakayosi [2], both of which were mainly developed for the character recognition researches. Since wide variety is observed in handwritten characters, for example, the size, slant, distortion, cursive or printed hand, tremor of pen movement in a stroke, it is necessary to collect samples from a large number of people for the purpose of character recognition. As for forensic handwriting identification, intra-individual difference or constancy along with inter-individual difference or variation is important. In a forensic handwriting identification case, a writer of a questioned handwriting can be known by comparing the questioned and the known handwriting. A document examiner cannot identify a writer if he does not know whether the difference derives from the intra-individual difference or the inter-individual difference. So, a dataset for forensic handwriting identification researches requires handwriting samples written by various writers with information on intra-individual constancy, that is, repetition. An offline dataset aimed for Japanese handwriting identification including sixty participants with 6-time repetitions [3] has been available to Japanese forensic document examiners, whose data were collected 1975 through 1976. The data have been, indeed, valuable but more than forty years have passed since the data collection. So, necessity for a new handwriting dataset for fundamental researches on forensic handwriting identification has been increasing.