Building Classifier Models for on-off Javanese Character Recognition

In this paper, we demostrated the building process of four classifier models as a part of an on-off character recognition system for Javanese characters. As Javanese character is no longer used in everyday writing and books, the dataset were collected by scanning the historical manuscripts and a reading lesson book. The rough dataset comprises 15.414 annotated characters and 633 classes. However, only 162 classes have sufficient data samples to be the training and testing one. Using this dataset, we measured the performance of four classifiers, namely k-NN, LDA, SVM, and Gaussian NB on the accuracy, micro-averaged precision, micro-averaged sensitivity and weighted-averaged precision and sensitivity metrices. The experiment shows that k-NN outperforms any other classifiers almost in most metrices, while SVM suffers the poorest performance. The research byproduct worth mentioning here is that it has identified 633 classes of distinct Javanese characters which comprise both common characters and compound characters found in modern Javanese writing as well as the archaic characters found in the literary works only.