hezarai/parsynth-ocr-4m

Synthetic Persian OCR Dataset (4M)

This dataset contains 4,000,000 synthetic Persian text line images generated using a customized Persian fork of TextRecognitionDataGenerator (trdg). It is designed specifically for training high-performance optical character recognition (OCR) and text recognition models for Persian text.

This dataset served as the foundational pre-training dataset for the official Hezar AI Persian OCR models, including hezarai/crnn-base-fa-v2 and hezarai/crnn-fa-license-plate-recognition-v2.

---

Dataset Overview

• Total Samples: 4,000,000 text line images
• Language: Persian (fa)
• Text Source: Sampled directly from the hezarai/lscp-pos-500k dictionary corpus.
• Generator Tools: hezarai/trdg-persian (featuring custom RTL fixes, ligature preservation, and word-split handling).
• Primary Use Case: Pre-training or fine-tuning OCR models (CRNN, TrOCR, Donut, etc.) for general Persian text and downstream fine-tuning (e.g., license plates, document scanning).

---

Generation Details & Splitting Strategy

The 4M dataset consists of three distinct partitions designed to expose OCR models to diverse text colors, background textures, and quality degradations:

Partition	Sample Count	Index Range	Background Type	Text Color Palette	Quality Degradation
Part 1 (Gaussian Noise)	3,000,000	0 – 2,999,999	Gaussian Noise (-b 0)	Dark tones (#284854, #542828)	50% downscaling, 20% applied (-rqf 0.5 -rqp 0.2)
Part 2 (Light Documents)	1,000,000	3,000,000 – 3,999,999	Light Doc Backgrounds (-b 3)	Dark tones (#284854, #542828)	Standard quality
Part 3 (Dark Documents)	1,000,000	4,000,000 – 4,999,999	Dark Doc Backgrounds (-b 3)	Soft pastel tones (#ed8c8c, #8ceced)	50% downscaling, 20% applied (-rqf 0.5 -rqp 0.2)

---

Generation Commands

The dataset was generated using the following three command executions:

# 1. Main subset: Gaussian noise backgrounds with quality degradation (3M)
trdg --output_dir ./data/fa-mixed -l fa -fd trdg/fonts/fa --dict trdg/dicts/fa_mixed_clean.txt \
  -t 48 -c 3000000 -sti 0 -w 1 -rqf 0.5 -rqp 0.2 -m 3,3,3,3 -tc "#284854,#542828" \
  -k 5 -rk -na 1 -ws -b 0 -f 48 -tdir rtl

# 2. Light document background subset (1M)
trdg --output_dir ./data/fa-mixed -l fa -fd trdg/fonts/fa --dict trdg/dicts/fa_mixed_clean.txt \
  -t 48 -c 1000000 -sti 3000000 -w 1 -m 3,3,3,3 -tc "#284854,#542828" \
  -k 5 -rk -na 1 -ws -b 0 -f 48 -tdir rtl -b 3 -id trdg/generators/documents_light

# 3. Dark document background subset with light text and quality degradation (1M)
trdg --output_dir ./data/fa-mixed -l fa -fd trdg/fonts/fa --dict trdg/dicts/fa_mixed_clean.txt \
  -t 48 -c 1000000 -sti 4000000 -w 1 -rqf 0.5 -rqp 0.2 -m 3,3,3,3 -tc "#ed8c8c,#8ceced" \
  -k 5 -rk -na 1 -ws -b 0 -f 48 -tdir rtl -b 3 -id trdg/generators/documents_dark

Parameter Breakdown

• -tdir rtl: Configures Right-to-Left rendering logic critical for Persian typography.
• -ws: Word-split mode enabled to preserve Persian ligatures and complex letter joining.
• -k 5 -rk: Applies a random rotation/skew angle between -5° and +5°.
• -m 3,3,3,3: Adds a 3-pixel margin surrounding the text.
• -f 48: Fixed image height set to 48 pixels.
• -fd trdg/fonts/fa: Utilizes a comprehensive directory of Persian TTF/OTF fonts.
• -na 1: Saved filename structure as [ID]_[TEXT].[EXT].

---

Pre-trained Models Trained on This Data

This dataset has proven effective for general Persian text recognition as well as downstream fine-tuning. Models trained on this dataset include:

• General Persian OCR:
  • hezarai/crnn-base-fa-v2
  • hezarai/crnn-base-fa-v1
• Specialized Fine-tuned Models:
  • hezarai/crnn-fa-license-plate-recognition-v2
  • hezarai/crnn-fa-license-plate-recognition-v1

---

Citation & Acknowledgments

If you use this dataset or the generation pipeline in your research or applications, please reference the Hezar Project and the custom Persian generator repository.

@misc{trdg-persian,
  author = {Hezar AI},
  title = {TRDG Persian: Synthetic Text Generator for Persian OCR},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{[https://github.com/hezarai/trdg-persian](https://github.com/hezarai/trdg-persian)}}
}