Handwritten Digit Recognition using CNN and ANN

Project Overview

This project uses Deep Learning to classify handwritten digits from the MNIST dataset. The MNIST dataset contains 28x28 grayscale images of digits from 0 to 9. We explore two different approaches for classification: Convolutional Neural Networks (CNNs) and Artificial Neural Networks (ANNs). The goal is to effectively predict the digit in each image and compare the performance of both models.

Key Features

Data Preprocessing

• Normalization: Pixel values were normalized using MinMaxScaler to bring them to the range [0, 1] for both CNN and ANN models.
• Reshaping: Reshaped images into tensors for CNN compatibility and into 1D vectors for ANN models.
• One-hot Encoding: Converted labels into one-hot encoded format using to_categorical to facilitate multi-class classification.

Exploratory Data Analysis (EDA)

• Visualization: Displayed sample images from the dataset to understand the structure and properties.
• Pixel Range Verification: Verified pixel values before and after normalization for scaling correctness.

CNN Model Development

• Convolutional Layers: Added multiple layers with ReLU activation and MaxPooling for feature extraction.
• Data Augmentation: Applied ImageDataGenerator for data augmentation to enhance generalization by shifting, rotating, and zooming images.
• Compilation and Optimization: Used Adam Optimizer with a learning rate schedule and Early Stopping for better convergence.

ANN Model Development

• Multiple Architectures: Built three different ANN models of varying complexity:
  • Model_1: Balanced architecture with moderate complexity.
  • Model_2: Simpler architecture with fewer layers and parameters.
  • Model_3: Deeper architecture with more layers and higher capacity to handle complexity.
• Model Selection: Model_1 was found to be the most suitable for digit classification, balancing complexity and performance.

Model Evaluation

• Metrics: Used accuracy, F1 score, confusion matrix, and classification report for evaluating model performance.
• Plots: Displayed training and validation accuracy/loss for all models to visualize learning progress and overfitting behavior.

Dependencies

• Programming Language: Python 3.x
• Libraries:
  • Data Manipulation: pandas, numpy
  • Visualization: matplotlib, seaborn
  • Deep Learning: tensorflow, keras
  • Preprocessing: scikit-learn