Hindi OCR API

An easy‑to‑use OCR service for handwritten Hindi text. Built with FastAPI, OpenCV, and TensorFlow/PyTorch, it detects word regions, extracts text, and classifies snippets.

---

🚀 Features

• Word Detection: Highlights words in uploaded images
• Text Extraction: Uses your custom OCR script to extract Hindi text (Devanagari)
• Model Prediction: Classifies text snippets via your .pth or .keras model
• Safe Temp‑File Handling: Works on Windows & Linux

---

📂 Repository Structure

hindi-ocr/
├── app.py                                  # FastAPI application  
├── requirements.txt                        # Python dependencies  
├── models/                                 # Model weight files  
│   │───notebooks/                          # Jupyter notebooks for trained model  
│   │   │── handwritten[pytorch].ipynb      # model train using pytorch
│   │   └── handwritten[tensorflow].ipynb   # model train using tensorflow
│   ├── hindi_ocr_model.pth                 # PyTorch model  
│   └── hindi_ocr_model.keras               # (optional TF fallback)  
├── fonts/                                  # Font files  
│   └── NotoSansDevanagari-Regular.ttf  
├── dataset/                                # Test data  
│   ├── images/                             # Input images for OCR  
│   │   └── training images                 # sample handwritten Hindi image  
│   └── words/                              # Expected‑output text files  
│       └── output labels                   # sample transcription  
├── label_encoder.pkl                       # sklearn LabelEncoder for class decoding 
│── Sample_OCR_Image.png                    # example image of OCR performace 
└── README.md                               # Project documentation  

---

Table of Contents

• Installation
• Configuration
• Running the API
• API Endpoints
• Example Usage
• Troubleshooting
• License

---

Installation

1. Clone repo
   git clone https://github.com/Stu-ops/hindi-ocr.git
cd hindi-ocr
2. Virtual environment
   python -m venv venv
source venv/bin/activate # macOS/Linux
venv\Scripts\activate # Windows
3. Install dependencies
   pip install -r requirements.txt
pip install python-multipart # for file uploads

---

Configuration

Ensure the following files/folders sit next to app.py:

• models/ (your .keras and optional .pth weights)
• label_encoder.pkl (sklearn LabelEncoder)
• fonts/NotoSansDevanagari-Regular.ttf
• dataset/ (training and testing images, words for testing)

---

Running the API

uvicorn app:app --reload --host 0.0.0.0 --port 8000

Swagger UI: /docs
ReDoc: /redoc

---

API Endpoints

Method	Path	Description
GET	/	Welcome HTML page
POST	/process/	Upload image → returns OCR & prediction
GET	/word-detection/	Returns word‑boxed image
GET	/prediction/	Returns prediction‑overlay image

Response Schema for /process/

{
  "OCR_output": "यह एक उदाहरण है",
  "word_count": 5,
  "prediction_label": "अ"
}
  

---

Example Usage

cURL

curl -X POST "http://localhost:8000/process/" \
  -H "Content-Type: multipart/form-data" \
  -F "file=@dataset/example1.png"

Python

import requests
url = "http://localhost:8000/process/"
with open("dataset/example1.png","rb") as f:
files = {"file": f}
resp = requests.post(url, files=files)
print(resp.json())

Retrieve word‑detection image:

curl http://localhost:8000/word-detection/ --output words.png

---

Troubleshooting

• “Form data requires python-multipart” → pip install python-multipart
• PermissionError on Windows → ensure you close image files before deletion; use with Image.open(...)
• Missing glyph warnings → add fallback font:
  plt.rcParams['font.family'] = ['Noto Sans Devanagari', 'DejaVu Sans']

License

MIT © Stu-ops