tell2jyoti/odia-handwritten-ocr

Odia Handwritten OCR Dataset

Dataset Description

This dataset contains 182,152 handwritten Odia character images prepared for training OCR models. The dataset covers all 47 OHCS (Odia Handwritten Character Set) characters with balanced class distribution.

Dataset Summary

• Language: Odia (ଓଡ଼ିଆ)
• Task: Optical Character Recognition (OCR)
• Total Images: 182,152
• Character Classes: 47
• Image Format: Grayscale JPG (32x32 pixels)
• Splits: Train (145,717), Validation (18,211), Test (18,224)

Supported Tasks

• Handwritten Character Recognition: Recognize individual Odia characters from images
• OCR Model Training: Train and fine-tune OCR models for Odia script
• Transfer Learning: Pre-train vision models on Odia characters

Languages

• Odia (ISO 639-1: or)

---

Dataset Structure

Data Instances

Each instance contains:

• image_path: Relative path to the image file
• absolute_path: Full system path (for reference)
• filename: Image filename
• class_id: Numeric class identifier (0-46)
• character: Odia character (Unicode)
• type: Image type ('simple', 'compound', 'synthetic')
• augmentation: Augmentation variant applied

Example:

{
  "image_path": "Odia_handwritten/.../0000010000.jpg",
  "absolute_path": "/full/path/to/image.jpg",
  "filename": "0000010000_resized.jpg",
  "class_id": 1,
  "character": "ଆ",
  "type": "simple",
  "augmentation": "grayscale"
}

Data Fields

• image_path (string): Relative path to image
• class_id (int): Character class ID (0-46)
• character (string): Odia Unicode character
• type (string): 'simple', 'compound', or 'synthetic'
• augmentation (string): Augmentation type applied

Data Splits

Split	Images	Purpose
Train	145,717	Model training
Validation	18,211	Hyperparameter tuning
Test	18,224	Final evaluation
Total	182,152	Complete dataset

---

Dataset Creation

Curation Rationale

This dataset was created to support research and development of OCR systems for the Odia language, which is spoken by ~40 million people but lacks comprehensive handwriting recognition resources.

Source Data

Initial Data Collection

• Handwritten Characters: 253,627 images from public Odia handwriting dataset
• Synthetic Characters: 51,000 images generated using Odia fonts with handwriting effects
• Augmented Images: 23,354 images created through data augmentation

Data Processing

1. Annotation: Mapped image filenames to Odia character labels
2. Balancing: Downsampled overrepresented classes and augmented underrepresented ones
3. Synthetic Generation: Created realistic handwritten-style images for missing character classes
4. Splitting: Stratified 80/10/10 split ensuring all classes in each split

Annotations

Annotations were created automatically by:

1. Extracting class IDs from image filenames
2. Mapping to Odia Unicode characters using odia_characters_label.csv
3. Validating character coverage across all 47 OHCS characters

---

Character Set

All 47 OHCS Odia Characters

Vowels (12): ଅ, ଆ, ଇ, ଈ, ଉ, ଊ, ଋ, ୠ, ଏ, ଐ, ଓ, ଔ

Consonants (33): କ, ଖ, ଗ, ଘ, ଙ, ଚ, ଛ, ଜ, ଝ, ଞ, ଟ, ଠ, ଡ, ଢ, ଣ, ତ, ଥ, ଦ, ଧ, ନ, ପ, ଫ, ବ, ଭ, ମ, ଯ, ର, ଲ, ଳ, ଶ, ଷ, ସ, ହ

Special (2): କ୍ଷ, ୟ

Class Distribution

All classes have between 3,000 to 10,000 samples for balanced training.

---

Usage

Loading the Dataset

from datasets import load_dataset
import json

# Load splits
train_data = load_dataset("tell2jyoti/odia-handwritten-ocr", split="train")
val_data = load_dataset("tell2jyoti/odia-handwritten-ocr", split="validation")
test_data = load_dataset("tell2jyoti/odia-handwritten-ocr", split="test")

# Or load from JSON directly
with open("train.json", "r") as f:
    train = json.load(f)

print(f"Training samples: {len(train['annotations']):,}")
print(f"Classes: {train['metadata']['unique_classes']}")

PyTorch DataLoader

from torch.utils.data import Dataset, DataLoader
from PIL import Image
import json

class OdiaOCRDataset(Dataset):
    def __init__(self, json_path, transform=None):
        with open(json_path, 'r') as f:
            data = json.load(f)
        self.annotations = data['annotations']
        self.transform = transform

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

    def __getitem__(self, idx):
        ann = self.annotations[idx]
        image = Image.open(ann['absolute_path']).convert('L')
        label = ann['class_id']

        if self.transform:
            image = self.transform(image)

        return image, label

# Create DataLoader
dataset = OdiaOCRDataset('train.json')
loader = DataLoader(dataset, batch_size=32, shuffle=True)

---

Additional Information

Dataset Curators

• Created by: tell2jyoti
• Project: DocuExtract - Odia Handwritten OCR
• GitHub: https://github.com/tell2jyoti/docuextract-odia-ocr

Licensing Information

This dataset is released under the MIT License.

Citation Information

@dataset{odia_handwritten_ocr_2026,
  title={Odia Handwritten OCR Dataset},
  author={Jyoti},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/tell2jyoti/odia-handwritten-ocr}}
}

Contributions

Thanks to the creators of the original Odia handwritten character dataset and the AI4Bharat team for IndicNLP resources.

---

Dataset Statistics

Image Distribution

Type	Images	Percentage
Original Handwritten	131,152	72%
Synthetic Generated	51,000	28%

Quality Metrics

• ✅ All 47 OHCS characters present
• ✅ Minimum 3,000 samples per class
• ✅ Balanced class distribution (3K-10K per class)
• ✅ Stratified splits (no overlap)
• ✅ Comprehensive metadata

---

Last Updated: January 2026 Version: 1.0.0 Status: Phase 1 Complete