Handwritten Digit Recognition using Deep Learning

This project implements a handwritten digit recognition system using the MNIST dataset. It showcases multiple deep learning techniques such as Dense Neural Networks, Convolutional Neural Networks (CNNs), Data Augmentation, Dropout, and Autoencoders for performance enhancement and image reconstruction.

🧠 Project Highlights

• Dataset: MNIST handwritten digits (28x28 grayscale images)
• Framework: TensorFlow / Keras
• Techniques:
  • Dense Neural Networks
  • Convolutional Neural Networks (CNNs)
  • Pooling (MaxPooling and AveragePooling)
  • Dropout regularization
  • Data Augmentation (horizontal flip, zoom)
  • Deeper CNN Architectures
  • Autoencoder for super-resolution reconstruction

📈 Final Results

Model	Test Accuracy
Dense Neural Network	~96.2%
Simple CNN (Conv + Flatten)	~97.7%
CNN + Average Pooling	~98.1%
CNN + Dropout Regularization	~98.6%
Deeper CNN (3 Conv layers)	~99.1%
Autoencoder Image Reconstruction	High-quality (loss ~0.0045)

---

---

📊 Key Features

• Model training with different architectures

• Evaluation using accuracy, loss curves

• Data Augmentation (zoom, horizontal flip)

• Dropout for regularization

---