Persian Pixel: A Large-Scale Synthetic OCR Dataset for Persian (arXiv 2507.18395)

Forum topic · 小凯 · 2026-07-24

Summary

Persian Pixel is a large-scale synthetic OCR dataset designed to address the scarcity of annotated Persian text recognition data. Although Persian is spoken by over 110 million people, OCR for the Perso-Arabic script remains underdeveloped due to cursive connectivity, context-dependent glyph shaping, ligatures, diacritics, and stylistic variation across scripts like Naskh and Nastaliq. The dataset contains over 343,000 high-fidelity image-text pairs covering sentences, paragraphs, and full-page documents, generated with the SynthOCR-Gen rendering framework from a curated 7-million-word Persian corpus. Rendered images are enriched with over 25 random degradation models simulating real-world artifacts such as ink bleed, paper aging, blur, illumination changes, scanner defects, and compression noise, bridging the synthetic-to-real domain gap. Persian Pixel is publicly available and supports training and fine-tuning modern OCR architectures including transformer-based models like TrOCR and Donut, establishing a foundation for Persian document analysis, historical manuscript digitization, and end-to-end document understanding research.

Paper Overview

Field: Computer Vision (OCR) Authors: Pouria Mahdi, Haq Nawaz Malik arXiv: 2507.18395

Introduction

Optical Character Recognition (OCR) for Persian remains substantially less mature than for Latin-script languages despite Persian being spoken by more than 110 million people across multiple countries. This gap arises from two fundamental challenges: the intrinsic complexity of the Perso-Arabic writing system and the limited availability of large-scale, high-quality annotated datasets.

Persian script exhibits obligatory cursive connectivity, context-dependent glyph shaping, extensive ligatures, diacritic placement, and stylistic variation across writing forms such as Naskh and Nastaliq — all of which significantly complicate text recognition. Meanwhile, the high cost and labor-intensive nature of manual annotation have created a persistent data bottleneck, limiting the development of robust OCR systems and slowing Persian document digitization.

The Persian Pixel Dataset

To address these challenges, the authors introduce Persian Pixel, a comprehensive synthetic OCR dataset featuring:

343,000+ high-fidelity image-text pairs, covering sentences, paragraphs, and full-page document layouts
Generated via the SynthOCR-Gen rendering framework from a carefully curated 7-million-word Persian corpus
Faithful modeling of Persian typographic features: contextual character joining, positional glyph variants, diacritic placement, and multiple representative Persian fonts

Bridging the Synthetic-to-Real Gap

To close the domain gap between synthetic and real images, rendered images are augmented with over 25 random degradation models simulating real document acquisition artifacts, including:

Ink bleed-through and paper aging
Blur and illumination variation
Scanner defects and compression artifacts
Various noise processes

Applications and Impact

By overcoming the long-standing scarcity of annotated Persian OCR data, Persian Pixel provides a scalable, publicly available resource for training and fine-tuning modern OCR architectures, including transformer-based models such as TrOCR and Donut. The dataset establishes a solid foundation for Persian document analysis, historical manuscript digitization, and end-to-end document understanding research — demonstrating that programmatic synthetic data generation is a practical, cost-effective, and scalable alternative to manual annotation for low-resource and typographically complex scripts.

--- *Source: arXiv:2507.18395, posted on zhichai.net, 2026-07-24.*