# Kannada OCR Model (443 Classes)

A deep learning model for Kannada character recognition using transfer learning with VGG16.

## Model Details

- **Architecture**: VGG16-based transfer learning
- **Classes**: 443 unique Kannada characters
- **Input Size**: 128x128 RGB images
- **Training Images**: 4,642 samples
- **Validation Images**: 905 samples

## Performance

- **Validation Accuracy**: 85.16%
- **Top-5 Accuracy**: 98.21%
- **Training Time**: ~3 hours 9 minutes on dual RTX 4090s

## Model Architecture

```
VGG16 (frozen weights)
├── BatchNormalization
├── Flatten
├── Dense(1024, relu) + L2 regularization
├── BatchNormalization + Dropout(0.5)
├── Dense(512, relu) + L2 regularization
├── BatchNormalization + Dropout(0.4)
└── Dense(443, softmax)
```

## Usage

### Download Model from Hugging Face

```python
from huggingface_hub import hf_hub_download
import tensorflow as tf
import pandas as pd

# Download model and class names
model_path = hf_hub_download(repo_id="srikarthikv/kannada-ocr-443-classes", filename="kannada_model_full.h5")
class_names_path = hf_hub_download(repo_id="srikarthikv/kannada-ocr-443-classes", filename="class_names.csv")

# Load model
model = tf.keras.models.load_model(model_path)

# Load class names
class_df = pd.read_csv(class_names_path)
class_names = class_df['class_name'].tolist()
```

### Prediction

```python
import cv2
import numpy as np

def preprocess_image(image_path):
    # Read and resize image
    img = cv2.imread(image_path)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = cv2.resize(img, (128, 128))

    # Normalize and add batch dimension
    img = img.astype(np.float32) / 255.0
    img = np.expand_dims(img, axis=0)

    return img

# Make prediction
image = preprocess_image("path/to/kannada_character.jpg")
predictions = model.predict(image)

# Get top prediction
top_class_idx = np.argmax(predictions[0])
confidence = predictions[0][top_class_idx]
predicted_character = class_names[top_class_idx]

print(f"Predicted Character: {predicted_character}")
print(f"Confidence: {confidence:.4f}")
```

### Interactive UI

Run the Streamlit UI for testing:

```bash
pip install streamlit opencv-python
streamlit run test_model_ui.py
```

## Files

- `kannada_model_full.h5` - Complete trained model (168 MB)
- `class_names.csv` - Character class mappings
- `test_model_ui.py` - Interactive Streamlit UI for testing

## Training Details

- **Framework**: TensorFlow/Keras
- **Base Model**: VGG16 (ImageNet weights, frozen)
- **Optimizer**: Adam (lr=0.001)
- **Loss**: Categorical Crossentropy
- **Batch Size**: 64
- **Epochs**: 60
- **Data Augmentation**: Rotation, shift, shear, zoom, brightness
- **Class Weights**: Balanced for imbalanced data

## Model Card

| Metric | Value |
|--------|-------|
| Model Size | 168 MB |
| Parameters | 17.3M total (2.6M trainable) |
| Training Time | 3h 9m |
| GPU Memory | ~12GB (dual RTX 4090) |
| Inference Time | ~10ms per image |

## Hugging Face Repository

🤗 **Model**: [srikarthikv/kannada-ocr-443-classes](https://huggingface.co/srikarthikv/kannada-ocr-443-classes)

## License

This model is released under MIT License for research and educational purposes.

## Citation

```bibtex
@misc{kannada-ocr-443,
  title={Kannada OCR Model with 443 Character Classes},
  author={Your Name},
  year={2025},
  url={https://huggingface.co/srikarthikv/kannada-ocr-443-classes}
}
```