Koshur-OCR-Synth

A large-scale synthetic optical character recognition (OCR) dataset for Kashmiri (ks). It consists of clean, high-resolution text-line images rendered programmatically in native Kashmiri Nastaliq and Naskh fonts, each paired with its exact ground-truth Unicode transcription.

Designed for OCR recognition, image-to-text modeling (e.g., TrOCR, PARSeq, Donut), fine-tuning, and benchmarking workflows that need clean, controllable image/label pairs at scale.

---

Sample Images

---

Dataset Summary

Property	Value
Language	Kashmiri (ks)
Script	Perso-Arabic (Nastaliq / Naskh)
Modality	Image → Text (OCR)
Source	Synthetic (programmatically font-rendered)
Fonts	Gulmarg Nastaleeq, Noto Nastaliq Urdu, Afan Koshur Naksh, Narqalam
Total Rows	285,864
Format	ZIP archive of PNG images and companion TXT labels
License	Apache 2.0

---

Dataset Structure

The dataset is packaged as a single ZIP file: Kashmiri_OCR_Nastaliq_Naskh_Dataset.zip (2.72 GB).

Inside the archive, files are organized as flat pairs:

• archive/image_XXXX_YYYYY.png: Rendered line image (RGB, height normalized/padding-ready).
• archive/image_XXXX_YYYYY.txt: Companion UTF-8 text label (exact transcription).

Fonts and Distribution

Font	Style	Share
Gulmarg Nastaleeq	Nastaliq	33.3%
Noto Nastaliq Urdu	Nastaliq	33.3%
Afan Koshur Naksh	Naskh	16.7%
Narqalam	Naskh	16.7%

---

Usage

Download and Extract using Python

from huggingface_hub import hf_hub_download
import zipfile, os

zip_path = hf_hub_download(
    repo_id="Faizaniqbal/Koshur-OCR-Synth",
    filename="Kashmiri_OCR_Nastaliq_Naskh_Dataset.zip",
    repo_type="dataset"
)

extract_dir = "./koshur_ocr_synth"
os.makedirs(extract_dir, exist_ok=True)
with zipfile.ZipFile(zip_path, 'r') as zip_ref:
    zip_ref.extractall(extract_dir)

print(f"Extracted to: {extract_dir}")

Loading in PyTorch

import os
from PIL import Image
from torch.utils.data import Dataset

class KoshurOCRDataset(Dataset):
    def __init__(self, data_dir, processor=None):
        self.data_dir = os.path.join(data_dir, "archive")
        self.basenames = sorted(set(
            f.rsplit(".", 1)[0] for f in os.listdir(self.data_dir) if f.endswith(".png")
        ))
        self.processor = processor

    def __len__(self):
        return len(self.basenames)

    def __getitem__(self, idx):
        base = self.basenames[idx]
        image = Image.open(os.path.join(self.data_dir, f"{base}.png")).convert("RGB")
        with open(os.path.join(self.data_dir, f"{base}.txt"), encoding="utf-8") as f:
            text = f.read().strip()
        if self.processor:
            px = self.processor(images=image, return_tensors="pt").pixel_values
            return {"pixel_values": px.squeeze(), "text": text}
        return {"image": image, "text": text}

---

Recommended Use Cases

1. Fine-tuning OCR/Image-to-Text Models: TrOCR, Donut, BLIP-2, PARSeq, PaliGemma for reading Kashmiri Nastaliq scripts.
2. Pretraining Stages: Use as a clean synthetic stage before introducing noisy real-world scans.
3. Benchmarking: Evaluate Kashmiri diacritic parsing (ٲ, ہ, گ, ۆ) and normalization pipelines under controlled conditions.

---

Limitations

• Synthetic only: Font-rendered performance does not guarantee parity with real scanned documents or handwriting.
• Font Diversity: Four fonts are represented; real-world print has wider stylistic variation.

---

License

Released under the Apache 2.0 License. Free to use, modify, and distribute for commercial or non-commercial purposes.