Optical character recognition: A literature review of challenges, methods, and future directions in the nusantara context

Abstract

Optical Character Recognition (OCR) has evolved from rule-based template matching to a multimodal sub-field of computer vision using deep neural networks and vision-language models. While modern OCR systems achieve near-human accuracy on printed Latin text, they struggle with low-resource scripts, degraded media, and writing systems with different visual logic. This review focuses on OCR in the Nusantara context, where over 700 languages coexist with ten traditional indigenous scripts and Jawi. It traces the methodological trajectory from feature-based classifiers to Convolutional Recurrent Neural Networks (CRNNs) with Connectionist Temporal Classification (CTC) loss, encoder-decoder transformers like TrOCR and Donut, and OCR-free vision-language paradigms. The review synthesises script-specific OCR research on Aksara Jawa, Aksara Bali, Sundanese, Batak, Lampung, Bugis-Makassar, and Jawi. Drawing on the NusaAksara benchmark and recent evaluations, the review argues that the challenge is data-related: the lack of curated datasets and the mismatch between abugida grammar and tokenisers for alphabetic languages. The paper concludes by proposing research directions that combine synthetic data generation, transfer learning from related Brahmic scripts, community-driven annotation, and culturally aware multimodal foundation models.