Tamil-Heritage-Ai

🏛️ Tamil Heritage AI

A research-grade Python application designed to digitize, segment, and recognize ancient Tamil script from degraded stone inscriptions and palm-leaf manuscripts. This project leverages advanced image preprocessing techniques and a custom-trained Convolutional Neural Network (CNN) to achieve high-accuracy character recognition, complete with explainable AI (Grad-CAM) visualization.

🌟 Key Features

• Advanced Image Preprocessing: Combines Non-Local Means (NLM) denoising, Contrast Limited Adaptive Histogram Equalization (CLAHE), and Adaptive Thresholding to clean heavily degraded heritage images.
• Smart Character Segmentation: Uses connected-component analysis with tight bounding box filtering to ignore stone textures and isolate valid characters in their natural reading order.
• Deep Learning Inference: Powered by a custom 3-block PyTorch CNN (model_stone.pth), trained specifically on ancient Tamil character shapes (44 classes), achieving ~96% validation accuracy.
• Explainable AI (Grad-CAM): Generates heatmaps indicating exactly which parts of the character stroke the neural network focused on to make its prediction, crucial for archaeological transparency.
• Interactive Dashboard: A rich Streamlit UI that allows users to upload images, tweak preprocessing sliders in real-time, inspect individual characters in an interactive zoom panel, and export the transcribed text.
• Multi-Format Export: Download the recognized script and coordinates as JSON, Text, or PDF files.

🚀 Getting Started

Prerequisites

• Python 3.9+
• Tesseract OCR (Optional, for fallback recognition)

Installation

1. Clone the repository:

   git clone https://github.com/Ricthi/Ai-heritage-.git
   cd Ai-heritage-
   

2. Install dependencies:

   pip install -r requirements.txt
   

   (Ensure you have streamlit, torch, torchvision, opencv-python, numpy, pillow, pandas, and streamlit-drawable-canvas installed).

Running the Application

Launch the Streamlit dashboard:

cd tamil_heritage_ai/Model-Creation
streamlit run main_app.py

The app will open automatically in your browser at http://localhost:8501.

🧠 Training Your Own Model

If you have new labeled datasets of ancient characters, you can easily retrain the CNN.

1. Organize Labels: Place your class folders inside tamil_heritage_ai/Labels/.
2. Prepare Dataset: Run the dataset prep script to automatically split your images into train and val sets:

   python prepare_dataset.py
   

3. Train the Model: Run the training script to generate a new model_stone.pth:

   python train_stone_cnn.py
   

4. Use It: The Streamlit app will automatically detect model_stone.pth. Select it from the sidebar dropdown!

📁 Repository Structure

• tamil_heritage_ai/Labels/ — Raw dataset containing folders of labeled ancient characters.
• tamil_heritage_ai/Model-Creation/data/ — The split 70/30 training and validation datasets.
• tamil_heritage_ai/Model-Creation/train_stone_cnn.py — PyTorch training script for the CNN.
• tamil_heritage_ai/Model-Creation/prepare_dataset.py — Data aggregation and splitting script.
• tamil_heritage_ai/Model-Creation/main_app.py — The primary Streamlit frontend application.
• tamil_heritage_ai/Model-Creation/model_stone.pth — The compiled weights and label mappings for the neural network.

---

Built to preserve and digitize the rich epigraphical history of the Tamil language.