Persian Pixel: A large-scale synthetic OCR dataset for Persian language

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

The Persian Pixel dataset is a new, large-scale synthetic Optical Character Recognition (OCR) dataset designed to overcome the challenges of Persian language OCR, which lags behind Latin-script languages due to script complexity and data scarcity. Comprising over 343,000 high-fidelity image-text pairs, the dataset is generated from a curated seven-million-word Persian corpus using the SynthOCR-Gen rendering framework. It spans sentence, paragraph, and full-page document layouts, faithfully modeling Persian script characteristics like cursive connectivity, contextual glyph shaping, and diacritic placement across seven diverse typefaces, including Naskh and Nastaliq. To bridge the synthetic-to-real domain gap, images are enriched with more than twenty-five stochastic degradation models, emulating realistic document acquisition artifacts such as ink bleed, blur, and noise. This publicly available resource aims to advance research in Persian document analysis and digitization.

Key takeaway

For Machine Learning Engineers developing Persian OCR systems, the Persian Pixel dataset offers a critical resource to overcome data scarcity. You can use this large-scale synthetic dataset to fine-tune transformer-based models like TrOCR and Donut, significantly improving recognition accuracy for complex Perso-Arabic script. Consider implementing curriculum learning, starting with cleaner data and progressing to heavily degraded samples, to build more robust and generalizable OCR solutions for document digitization and historical archives.

Key insights

Synthetic data generation is a scalable, cost-effective solution for OCR in low-resource, complex scripts like Persian.

Principles

• OCR for complex scripts needs shaping-aware rendering.
• Diverse fonts and degradations improve model robustness.
• Multi-granularity data supports varied recognition tasks.

Method

The SynthOCR-Gen pipeline renders text from a curated corpus, applies contextual shaping, uses seven fonts, and adds 25+ stochastic degradations.

In practice

• Fine-tune transformer OCR models (TrOCR, Donut).
• Implement curriculum learning from clean to degraded data.
• Pre-train for domain adaptation with small real datasets.

Topics

• Optical Character Recognition
• Persian Language
• Synthetic Data
• Perso-Arabic Script
• Document Digitization
• TrOCR
• Donut

Code references

• Belval/TextRecognitionDataGenerator
• persiandataset/Arshasb

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Persian Pixel: A large-scale synthetic OCR dataset for Persian language — Synthetic data generation, exemplified by Persian Pixel, offers a scalable solution for OCR in typographically complex, low-resource languages.
• KazakhOCR: A Synthetic Benchmark for Evaluating Multimodal Models in Low-Resource Kazakh Script OCR — Current MLLMs struggle with OCR and language identification for low-resource Kazakh scripts, especially Arabic, performing worse than traditional OCR.
• Advancing WordArt-Oriented Scene Text Recognition: Datasets and Methods — Recognizing highly customized WordArt demands specialized datasets and models beyond general Scene Text Recognition.
• LV-ROVER-MLT: Low-Resource Maltese OCR by Synthetic Fine-Tuning and Multi-Stream Arbitration — Synthetic data and multi-stream ensembles significantly improve low-resource OCR, especially for diacritic-rich languages.
• Why Nepali OCR Is Brutally Hard — And How We Finally Solved It — A hybrid deep learning pipeline overcomes complex script challenges to automate document data extraction and form pre-filling.
• Urdu Katib Handwritten Dataset: A Historical Document Dataset for Offline Urdu Handwritten Text Recognition with CRNN-Based Baseline Evaluation — The Urdu Katib Handwritten Dataset (UKHD) and CRNN-BGRU-CTC model establish a baseline for historical Urdu handwritten text recognition.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.