AliShafiee2003/persian-ocr-garshasp-70c

Persian OCR Garshasp: Large-Scale Synthetic Persian OCR Dataset

Persian-OCR-Garshasp is a large-scale Persian (Farsi) OCR dataset containing about 2.6M image–text pairs. It is designed for training and evaluating Optical Character Recognition (OCR) and scene text recognition models for Persian text in the wild.

Each sample consists of:

• an RGB image with height 48 px and width 640 px
• up to 70 Persian (Farsi) characters as the ground-truth text label
• a style field describing the rendering / distortion style (e.g. plain_text, sine_text, parabolic_text, white_noise, *_blur, *_shadow, *_darkness, plain_text_complex_bg, noisy_Bg)

The dataset is fully synthetic: all images are generated programmatically from Persian text using multiple fonts, styles, geometric distortions and noisy backgrounds in order to approximate real-world Persian OCR scenarios (screenshots, signs, documents, noisy photos, etc.) as closely as possible.

---

Dataset summary

• Name: Persian OCR Garshasp
• Author: AliShafiee2003
• Language: Persian (Farsi)
• Script: Persian (Arabic-based) script
• Primary task: Image-to-Text / Optical Character Recognition (OCR)
• Secondary tasks:
  • Scene text recognition
  • Document text recognition
• Modality: Image → Text
• Total size: ~2.6M image–text pairs (14 splits)
• Image resolution: 48 × 640 (height × width), RGB
• Max text length: 70 characters per image
• Number of styles: 13 rendering / distortion styles
• Recommended usages:
  • Training Persian / Farsi OCR models
  • Pretraining on synthetic data before fine-tuning on real-world images
  • Robustness experiments to geometric distortions, blur, noise and complex backgrounds

---

Data fields

Each row has the following fields:

Field	Type	Description
image	Image	RGB image, height 48 px, width 640 px, containing Persian text
text	string	Ground-truth Persian (Farsi) text (up to 70 characters)
style	string	Rendering / distortion style (one of 13 possible values)

Example style values (non-exhaustive):

• plain_text
• plain_text_blur
• plain_text_shadow
• plain_text_darkness
• plain_text_complex_bg
• sine_text
• sine_text_blur
• sine_text_darkness
• parabolic_text
• parabolic_text_blur
• parabolic_text_shadow
• white_noise
• noisy_Bg

---

Splits

The dataset is organized into one large split plus several style-specific splits:

• train – 1.3M rows (main training split)
• noisy_Bg – 100k rows
• parabolic_text – 100k rows
• parabolic_text_blur – 100k rows
• parabolic_text_shadow – 100k rows
• plain_text – 100k rows
• plain_text_blur – 100k rows
• plain_text_complex_bg – 100k rows
• plain_text_darkness – 100k rows
• plain_text_shadow – 100k rows
• sine_text – 100k rows
• sine_text_blur – 100k rows
• sine_text_darkness – 100k rows
• white_noise – 100k rows

You can:

• use only the train split for large-scale training, or
• build your own mixtures from style-specific splits to control the distribution of distortions / backgrounds.

---

Data source & generation

Text

• Text length:
  • Each sample contains at most 70 characters.
• Preprocessing (recommended, if applicable):
  • Normalization of Persian characters (e.g. unify ی/ي and ک/ك, remove zero-width characters, etc.).
  • Removal of control characters and very short / empty strings.
  • Optional filtering of non-Persian content (numbers, English tokens, etc.).

Please fill in the exact details of your preprocessing so others can better understand and reproduce the pipeline.

Fonts

The Persian text is rendered using 82 fonts from 18 font families, including:

1. Arial
2. B Homa
3. B Lotus
4. B Mitra
5. B Tir
6. B Traffic
7. B Yagut
8. B Zar
9. B Nazanin
10. Calibri
11. Dubai
12. Estedad
13. Far Farnaz
14. IRANSans
15. Sahel
16. Shabnam
17. Tahoma
18. Vazir / Vazirmatn

(Several weights and variants of these families are used to reach a total of 82 fonts.)

Rendering

• All images are rendered as RGB with fixed size 48×640 (height × width).

• Text is laid out to fit within the 640-pixel width, with at most 70 characters per image.

• Different styles control geometric distortion and background / noise:

  • Geometric distortions

    • sine_text: sinusoidal warping of text baselines
    • parabolic_text: parabolic / curved baselines and perspective-like distortion

  • Photometric / noise effects

    • Blur (*_blur)
    • Dark / low-light (*_darkness)
    • Shadows (*_shadow)
    • Complex backgrounds (plain_text_complex_bg)
    • White noise and noisy backgrounds (white_noise, noisy_Bg)

Synthetic nature

This dataset is fully synthetic:

• Persian text is sampled from a corpus and rendered with multiple fonts.
• Distortions and augmentations simulate conditions such as:
  • screenshots and UI text
  • signs, banners, boards
  • noisy / low-quality photos
  • documents with various backgrounds

Although synthetic, the pipeline is designed to approximate real-world Persian OCR as much as possible.

---

Usage

Load with 🤗 Datasets

from datasets import load_dataset

dataset = load_dataset("AliShafiee2003/persian-ocr-garshasp", split="train")

example = dataset[0]
image = example["image"]   # PIL.Image.Image (48x640, RGB)
text = example["text"]     # str (Persian)
style = example["style"]   # str

Streaming (recommended for full dataset)

from datasets import load_dataset

streaming_dataset = load_dataset(
    "AliShafiee2003/persian-ocr-garshasp",
    split="train",
    streaming=True,
)

for example in streaming_dataset:
    image = example["image"]
    text = example["text"]
    style = example["style"]
    # training loop here

PyTorch example

from datasets import load_dataset
from torch.utils.data import DataLoader
from torchvision import transforms

dataset = load_dataset(
    "AliShafiee2003/persian-ocr-garshasp",
    split="train",
)

transform = transforms.Compose([
    transforms.ToTensor(),
])

def collate_fn(batch):
    images = [transform(x["image"]) for x in batch]
    texts = [x["text"] for x in batch]
    styles = [x["style"] for x in batch]
    return {
        "pixel_values": images,
        "labels": texts,
        "styles": styles,
    }

dataloader = DataLoader(
    dataset,
    batch_size=32,
    shuffle=True,
    collate_fn=collate_fn,
)

---

Intended uses

Recommended use-cases:

• Training and evaluation of OCR models for Persian (Farsi).

• Pretraining scene text recognition models on synthetic data before fine-tuning on:

  • real scanned documents,
  • mobile photos of text,
  • screenshots / UI text.

• Robustness studies under:

  • geometric distortions (sine / parabolic),
  • blur and shadows,
  • noisy and complex backgrounds,
  • low-light conditions.

Not recommended for:

• Handwritten text recognition (this dataset focuses on printed text).
• Pure NLP / language modeling tasks (the text is primarily used as labels for images).

---

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

You are free to:

• Share — copy and redistribute the material in any medium or format
• Adapt — remix, transform, and build upon the material for any purpose, even commercially

Under the following terms:

• Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use.

For more details, see the full license text:
https://creativecommons.org/licenses/by/4.0/

---

Citation

If you use Persian OCR Garshasp in your research or product, please cite:

@misc{persian_ocr_garshasp,
  title        = {Persian OCR Garshasp: Large-Scale Synthetic OCR Dataset for Persian (Farsi)},
  author       = {Shafiee, Ali},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/AliShafiee2003/persian-ocr-garshasp-70c}},
  note         = {Version 1.0},
}

(Feel free to adjust the author list, year, and version to match your preferred citation.)

---

خلاصه فارسی

Persian-OCR-Garshasp یک دیتاست بزرگ برای OCR فارسی است که حدود ۲.۶ میلیون جفت تصویر–متن دارد. هر نمونه شامل یک تصویر رنگی با سایز ۴۸×۶۴۰ پیکسل و متن فارسی (حداکثر ۷۰ کاراکتر) به‌عنوان برچسب است.

متن‌ها با استفاده از ۸۲ فونت از ۱۸ خانواده‌ی مختلف (مثل B Nazanin، B Zar، IRANSans، Sahel، Shabnam، Vazir/Vazirmatn و …) رندر شده‌اند و روی آن‌ها انواع اعوجاج‌های هندسی (sine، parabolic) و نویزها (تاری، سایه، پس‌زمینه پیچیده، نویز سفید و …) اعمال شده است.

این دیتاست کاملا مصنوعی است، اما سعی شده تا حد امکان به شرایط واقعی متن فارسی در تصاویر نزدیک باشد (اسکرین‌شات، تابلو، پوستر، عکس‌های نویزی، اسکن اسناد و غیره). از این دیتاست می‌توان برای آموزش و ارزیابی مدل‌های تشخیص متن فارسی، و همین‌طور پیش‌آموزش مدل‌ها قبل از فاین‌تیون روی دیتای واقعی استفاده کرد.

این دیتاست تحت لایسنس CC BY 4.0 منتشر شده است. یعنی.

---

Keywords (for search & discoverability)

persian ocr, farsi ocr, persian text recognition, farsi text recognition, persian scene text, persian document text, arabic script ocr, persian dataset, farsi dataset, synthetic ocr dataset, image to text, scene text recognition, document text recognition, persian news text, persian image text pairs, 48x640 images, noisy text images, sine text, parabolic text, white noise, shadow, blur, dark images, complex background, persian computer vision, persian deep learning