https://github.com/nithin1729s/kannada-cnn
This project recognizes 49 Kannada letters using a CNN classifier (TensorFlow/Keras). A FastAPI backend and Next.js (TypeScript) frontend enable users to draw letters and get predictions.
https://github.com/nithin1729s/kannada-cnn
cnn cnn-classification fastapi kannada kannada-mnist keras nextjs tensorflow typescript
Last synced: 3 months ago
JSON representation
This project recognizes 49 Kannada letters using a CNN classifier (TensorFlow/Keras). A FastAPI backend and Next.js (TypeScript) frontend enable users to draw letters and get predictions.
- Host: GitHub
- URL: https://github.com/nithin1729s/kannada-cnn
- Owner: Nithin1729S
- Created: 2025-03-12T19:07:22.000Z (over 1 year ago)
- Default Branch: master
- Last Pushed: 2025-04-13T02:21:58.000Z (about 1 year ago)
- Last Synced: 2025-06-17T07:47:55.082Z (about 1 year ago)
- Topics: cnn, cnn-classification, fastapi, kannada, kannada-mnist, keras, nextjs, tensorflow, typescript
- Language: Jupyter Notebook
- Homepage: https://kannada-cnn.vercel.app
- Size: 4.78 MB
- Stars: 2
- Watchers: 1
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Kannada Alphabet Recognition using CNN
This project implements a complete pipeline to recognize 49 Kannada letters. It combines deep learning for classification with a modern web application stack. The CNN classifier is built using TensorFlow/Keras, while a FastAPI backend and Next.js (TypeScript) frontend provide a interface where users can draw a letter and receive a prediction.
## Demo
[Screencast from 2025-03-13 01-00-34.webm](https://github.com/user-attachments/assets/2ac763ac-1f91-41d5-849c-864a2a288d87)
[Watch it on Youtube](https://youtu.be/YmIuzqwVocw)
## Overview
The project has two main components:
1. **Model Training & Evaluation:**
A convolutional neural network (CNN) is trained on a dataset of 49 Kannada alphabets. The training process involves data loading, augmentation, network construction, training with callbacks, and model evaluation.
2. **Web Application:**
- **FastAPI Backend:** Serves the trained model for real-time inference via API endpoints.
- **Next.js Frontend:** Provides a canvas for users to draw a letter. The drawn image is sent to the FastAPI backend, and the predicted output is displayed.
## Training Process
### Data Loading and Preprocessing
- **Dataset:**
The dataset can be obtained from [IEEE Dataport](https://ieee-dataport.org/documents/kannada-language-image-dataset).
- **Image Configuration:**
Images are converted to grayscale and resized to 28x28 pixels.
The dataset is split into training (80%) and validation (20%) sets using `tf.keras.utils.image_dataset_from_directory`, with labels one-hot encoded.
### Data Augmentation
To improve the generalization of the model, data augmentation is applied **only on the training set**:
- **Random Rotation**
- **Random Zoom**
- **Random Translation**
This augmentation is implemented using a Keras `Sequential` model.
### CNN Architecture
The CNN model consists of three convolutional blocks:
- **Block 1:**
- Two convolutional layers (32 filters, 3×3 kernel, ReLU activation) with batch normalization.
- Max pooling layer and dropout (25%).
- **Block 2:**
- Two convolutional layers (64 filters, 3x3 kernel, ReLU activation) with batch normalization.
- Max pooling and dropout (25%).
- **Block 3:**
- Two convolutional layers (128 filters, 3x3 kernel, ReLU activation) with batch normalization.
- Max pooling and dropout (25%).
Following the convolutional layers, the network is flattened and fed through dense layers:
- A dense layer with 128 neurons (ReLU activation) with batch normalization and dropout (50%).
- A final dense layer with a softmax activation that outputs probabilities for the 49 classes.
### Compilation and Training
- **Compilation:**
The model is compiled with the Adam optimizer and categorical crossentropy loss.
- **Callbacks:**
- **EarlyStopping:** Monitors the validation loss and stops training if no improvement is seen for 5 epochs.
- **ReduceLROnPlateau:** Reduces the learning rate when the validation loss plateaus.
- **Training:**
The model is set to train for up to 480 epochs; however, early stopping typically halts training sooner if the model converges.
- **Model Saving:**
Once training is complete, the model (both architecture and weights) is saved as `full_model.h5` for later use in inference.
## Inference and Evaluation
The trained model achieved 98.45% test accuracy.
## Setup and Installation
### Model Training
Model training can be found in Jupyter Notebook present in the notebooks folder.
### Environmental Variables
Create a `.env` file in the frontend directory with the following variables:
```env
NEXT_PUBLIC_BACKEND_URL=fastapi_address # URL of the FastAPI Backend Server
```
### FastAPI Backend
1. **Navigate to the `backend` Directory:**
```bash
cd backend
```
2. **Install Dependencies:**
```bash
pip install -r requirements.txt
```
3. **Run the FastAPI Server:**
```bash
uvicorn api.main:app --reload
```
### Next.js Frontend
1. **Navigate to the `frontend` Directory:**
```bash
cd frontend
```
2. **Install Dependencies:**
```bash
npm install
```
3. **Start the Development Server:**
```bash
npm run dev
```
4. **Open in Browser:**
Visit `http://localhost:3000` to interact with the application.