🧮 MNIST Digit Recognition Project

Welcome to the MNIST Digit Recognition project! This repository demonstrates a complete workflow for building, training, and deploying a neural network to recognize handwritten digits from the MNIST dataset. Additionally, it includes an interactive Streamlit web app for showcasing the model's capabilities.

https://www.kaggle.com/datasets/hojjatk/mnist-dataset

---

✨ Features

• 🧠 Model Training: Train a neural network to classify handwritten digits with high accuracy.
• 🎨 Interactive Drawing: Draw digits directly on a canvas for real-time classification.
• 📂 Image Upload: Upload digit images for prediction.
• 💾 Model Reusability: Save the trained model and reload it for predictions.
• 📊 Model Performance Metrics: Display detailed training and evaluation results.

---

🛠️ Technologies Used

• Python 3.10+: Programming language.
• TensorFlow/Keras: Deep learning framework for model building and training.
• Streamlit: Web app framework for creating the interactive user interface.
• NumPy: Numerical computing library.
• Matplotlib: Data visualization for debugging and display.
• OpenCV: Image processing for resizing and normalization.

---

🚀 Setup Instructions

1. Clone the repository:

   git clone https://github.com/alphatechlogics/HandWrittenDigitalsClassification.git
   cd mnist-digit-recognition

2. Install dependencies:

   pip install -r requirements.txt

3. Run the Streamlit app:

   streamlit run app.py

🧑‍💻 Model Training Workflow

• Data Preprocessing:

  • Normalize pixel values to [0, 1] for faster convergence.
  • Reshape input to match the neural network's requirements.

• Model Architecture:

  • A sequential neural network with:
    • Input layer to flatten the 28x28 grayscale image.
    • Fully connected dense layers with ReLU activation and dropout for regularization.
    • Output layer with softmax activation for class probabilities.

• Training:

  • Optimizer: Adam with a learning rate of 0.0003.
  • Loss Function: Sparse Categorical Crossentropy.
  • Metrics: Accuracy.

• Model Saving:

  • The trained model is saved in HDF5 format (model.h5).

🌟 Streamlit App Features

• Sidebar Features:

  • 🎥 Lottie Animation for visual engagement.
  • 📋 Information about the model and its features.
  • 📧 Contact section for support or inquiries.

• Interactive Prediction:

  • Choose between drawing a digit on a canvas or uploading an image.
  • Real-time predictions with confidence levels.
  • Easy-to-use buttons for classification.

📊 Results

• Training Accuracy: Achieved ~99% accuracy on the training dataset.
• Validation Accuracy: ~98% on the validation dataset.
• Test Accuracy: ~98.3% on the unseen test data.