Machine Learning — Image Classification

A collection of supervised machine learning models for binary image classification, implemented in Python using PyTorch and scikit-learn. Each notebook trains a different algorithm on the same dataset and evaluates it with a consistent set of metrics.

---

Task

Binary classification of images into two categories:

• 1 — Pleasant
• 0 — Unpleasant

Images are resized to 64×64 RGB. Each notebook independently loads the dataset, trains a model, evaluates on a validation set, and outputs test predictions to submission.csv.

---

Models Implemented

Notebook	Algorithm	Framework
CNN.ipynb	Convolutional Neural Network	PyTorch
MLP.ipynb	Multi-Layer Perceptron	PyTorch
SVM.ipynb	Support Vector Machine (grid search)	scikit-learn
EnsembleLearing.ipynb	Random Forest, Bagging SVM, AdaBoost	scikit-learn
kNN.ipynb	k-Nearest Neighbors (cosine distance)	scikit-learn
LogisticRegression.ipynb	Logistic Regression	scikit-learn

---

Model Details

CNN (CNN.ipynb)

• Architecture: 4 Conv layers (32→64→128→256 filters) + 2 FC layers
• Batch normalization + Dropout for regularization
• Optimizer: Adam (lr=0.0001), Loss: CrossEntropyLoss
• Training: 15 epochs

MLP (MLP.ipynb)

• Architecture: Input → [512, 256] → 2 output units
• Activation: PReLU (configurable: relu/sigmoid/tanh/prelu)
• Optimizer: Adam (lr=0.0001), Loss: CrossEntropyLoss
• Training: 20 epochs, batch size 64

SVM (SVM.ipynb)

• Grid search over 8 configurations: 2 kernels (linear, RBF) × 4 C values (0.1, 1, 10, 100)
• Best model selected by F1 score on validation set
• Comparative ROC curves for all configurations

Ensemble Methods (EnsembleLearing.ipynb)

• Random Forest: 3 configurations (50, 100, 200 trees, varying depths)
• Bagging + Linear SVM: 20 estimators
• AdaBoost: Decision Tree base, 200 estimators, max_depth=3

kNN (kNN.ipynb)

• k=3, cosine distance metric
• Flattened image vectors (64×64×3 = 12,288 features)

Logistic Regression (LogisticRegression.ipynb)

• C=0.5, solver=lbfgs, max_iter=1000
• Normalized input features to [0, 1]

---

Evaluation

Every model is evaluated with the same metrics on a held-out validation set (80/20 split, random_state=42):

• Accuracy, Precision, Recall, F1 Score
• Confusion Matrix
• ROC Curve (AUC)

---

Tech Stack

Library	Purpose
PyTorch	CNN and MLP implementation
scikit-learn	SVM, ensemble methods, kNN, logistic regression
OpenCV	Image loading and resizing
pandas	Data handling
matplotlib	Visualization (ROC curves, confusion matrices)

---

Project Structure

Machine-Learning/
├── CNN.ipynb                # Convolutional Neural Network
├── MLP.ipynb                # Multi-Layer Perceptron
├── SVM.ipynb                # Support Vector Machine (grid search)
├── EnsembleLearing.ipynb    # Random Forest, Bagging, AdaBoost
├── kNN.ipynb                # k-Nearest Neighbors
└── LogisticRegression.ipynb # Logistic Regression

---

Academic Context

Built as a university project for the Machine Learning course at the Department of Computer Engineering, University of Ioannina, Greece.

Key concepts implemented:

• Supervised binary classification
• Feature extraction from raw images (flattening, normalization)
• Deep learning with PyTorch (CNN, MLP)
• Classical ML with scikit-learn (SVM, kNN, Logistic Regression, Ensembles)
• Hyperparameter search and model comparison
• Evaluation metrics: F1, AUC-ROC, confusion matrix