A Baseline Evaluation of OCR Segmentation and Classification Methods for Printed Javanese Script

Optical Character Recognition (OCR) for Javanese script remains challenging due to its complex glyph structure and overlapping components. In this context, this study presents a pilot investigation into the segmentation and transliteration performance of printed Javanese text using a single-page dataset. The workflow begins with preprocessing, followed by segmentation and script-level classification using the k-Nearest Neighbor (k-NN) algorithm. Experiments were conducted with 91 and 281 scripts, utilizing 100 and 500 samples per script to evaluate system performance. As a result, the segmentation stage achieved an average accuracy of 63.5%, with some lines reaching above 80%. Moving to transliteration, accuracy comparisons were performed between segmentation output, model predictions, and ground truth, yielding average accuracies of 61.4%, 72.3%, and 39.2%, respectively. Further experiments using 3-NN and 11-NN configurations demonstrated that increasing training samples and script diversity improved recognition accuracy, achieving up to 68.75% on certain lines. This research provides an initial benchmark dataset and a systematic evaluation framework, establishing a baseline that bridges the gap between handwritten and printed OCR research. The findings offer empirical insights for developing robust OCR systems to support the digital preservation of Indonesia's written cultural heritage.