Introduction to Machine Learning - Digits Classification

This project is a comprehensive study and implementation of machine learning techniques applied to the classification of handwritten digits using the digits dataset from sklearn.

Project Overview

The goal is to explore, preprocess, and model the handwritten digits dataset to classify images of digits (0-9) with high accuracy.

Key steps include:

• Data loading and visualization
• Data normalization and dimensionality reduction (PCA)
• Feature engineering (zone-based and edge detection features)
• Building and evaluating models including Support Vector Machine (SVM) and Neural Networks
• Hyperparameter tuning and cross-validation
• Comparison of multi-class strategies (One-vs-Rest, One-vs-One)

Project Structure

• partie1Metier.py - Data loading, exploration, preprocessing, PCA and feature engineering
• partie2Metier.py - Dataset splitting, pipeline creation, SVM training, evaluation, hyperparameter tuning, and comparison of OvO/OvR strategies
• partie3Metier.py - Neural network model using Keras for digit classification
• RapportMachineLearning.pdf - Detailed project report including methodology, results, discussions, and conclusions

Requirements

• Python 3
• numpy
• matplotlib
• scikit-learn
• tensorflow (for neural network)
• scikit-image

How to run

1. Install dependencies (e.g., via pip): # Introduction to Machine Learning - Digits Classification

This project is a comprehensive study and implementation of machine learning techniques applied to the classification of handwritten digits using the digits dataset from sklearn.

Project Overview

The goal is to explore, preprocess, and model the handwritten digits dataset to classify images of digits (0-9) with high accuracy.

Key steps include:

• Data loading and visualization
• Data normalization and dimensionality reduction (PCA)
• Feature engineering (zone-based and edge detection features)
• Building and evaluating models including Support Vector Machine (SVM) and Neural Networks
• Hyperparameter tuning and cross-validation
• Comparison of multi-class strategies (One-vs-Rest, One-vs-One)

Project Structure

• partie1Metier.py - Data loading, exploration, preprocessing, PCA and feature engineering
• partie2Metier.py - Dataset splitting, pipeline creation, SVM training, evaluation, hyperparameter tuning, and comparison of OvO/OvR strategies
• partie3Metier.py - Neural network model using Keras for digit classification
• RapportMachineLearning.pdf - Detailed project report including methodology, results, discussions, and conclusions

Requirements

• Python 3
• numpy
• matplotlib
• scikit-learn
• tensorflow (for neural network)
• scikit-image

Refer to the PDF report for detailed explanations, results, and analysis.

---

Results Summary

• The best model achieved approximately 99% accuracy using an optimized SVM with RBF kernel, PCA, and combined handcrafted features (zones + Sobel edges).
• Neural networks showed promising results as well with clear training and validation loss convergence.
• Extensive hyperparameter tuning and comparison of multi-class strategies (OvR vs OvO) are included.

---

Authors

• Amiel Metier
• Hedi Chennoufi
• Lucas Fert

Supervised by:

• Jean-Luc Bouchot
• Mahmoud Elsawy