A Preliminary Study on Handwriting Classification Using Motor Information

Writing with a finger or a stylus pen on a tablet, especially in the case of signing, has been more and more popular in our ordinary life. The necessity of the analysis of handwriting data acquired online for forensic purposes has also been more and more important. There has been, however, a very small number of research conducted on online handwriting data analysis from the standpoint of forensic document examination. Handwriting classification experiments using the dataset collected both offline and online for forensic purposes were conducted in this paper. Three kinds of handwriting classification experiments using offline data and two kinds of online data respectively were conducted. The classification task was to use handwriting samples from 10 different writers to classify 10 different writers. A convolutional neural network algorithm was used for classification. In the experiments, four kinds of Japanese handwriting samples, which were acquired both online and offline, were combined to be a “signature” of a Japanese person and analyzed. Samples were divided into training samples and test samples and the classification experiments were done. After learning the handwriting features of the 10 writers using the training data, the test samples were classified into 10 writers. Online data used for the classification were pen tip trajectory with pen pressure or with writing speed. A trajectory was painted in three colors (pen pressure condition) or five colors (writing speed condition) like a map. Handwriting classification experiments were done on offline data, pen pressure data and writing speed data using LeNet respectively. The results showed over 95% accuracy in offline and pen pressure conditions. Writing speed data showed a lower accuracy score. The results suggested the availability of online data for forensic handwriting classification.