TreasureX Researchcs.AIPersian Pixel: A Massive Synthetic Dataset to Boost Persian OCR
Pouria Mahdi, Haq Nawaz Malik · July 22, 2026
Optical Character Recognition (OCR) for Persian lags far behind Latin languages, despite over 110 million speakers. The script's cursive nature, context-dependent letter shapes, and complex diacritics make it tough for machines to read. Worse, high-quality annotated data is scarce and expensive to produce.
Enter Persian Pixel, a synthetic dataset of over 343,000 image-text pairs. Researchers built it using a seven-million-word Persian corpus and a rendering engine that faithfully mimics Persian typography—including Naskh and Nastaliq styles. To make synthetic images look real, they added 25+ types of realistic degradation like ink bleed, blur, and paper aging.
This dataset allows modern OCR models (e.g., TrOCR, Donut) to train on realistic Persian documents without costly manual annotation. It's a scalable, open resource that could accelerate digitization of Persian books, historical manuscripts, and administrative records.
Why it matters: For investors, this opens up markets in Persian-speaking regions (Iran, Afghanistan, Tajikistan) for document automation, digital archiving, and AI-driven translation services. The synthetic data approach also sets a precedent for other complex scripts (Arabic, Urdu), potentially creating new niches in multilingual OCR.
Source: arXiv (https://arxiv.org/abs/2607.20385)
Educational content only, generated with AI assistance from public research abstracts. Not investment advice. Thank you to arXiv for use of its open access interoperability.