Car vs Bike Image Classification

Overview

This project explores the effectiveness of various machine learning approaches for distinguishing between car and motorcycle images. The project benchmarks four different algorithms with hyperparameter tuning and data augmentation.

Best Performance: CNN with data augmentation achieved 89.88% accuracy

Key Features

• Multiple Model Implementations: KNN, Random Forest, SVM, and CNN
• Data Augmentation: Horizontal flips, Gaussian blur, and noise injection
• Dimensionality Reduction: PCA analysis with varying component counts
• Comprehensive Evaluation: Accuracy, Precision, Recall, and F1-Score metrics
• Hyperparameter Optimization: Systematic tuning for each algorithm
• Comparative Analysis: Detailed performance comparison across all models

Dataset

• Source: Car vs Bike Classification Dataset (Kaggle)
• Size: 4,000 images (2,000 cars + 2,000 motorcycles)
• Preprocessing:
  • Resized to 64×64 pixels
  • Converted to grayscale
  • Flattened for classical ML algorithms
• Split: 60% training, 20% validation, 20% testing

Models & Results

Model	Configuration	Test Accuracy	Recall	Precision	F1-Score
CNN (Augmented)	4-layer architecture	89.88%	0.930	0.935	0.933
CNN (Non-augmented)	4-layer architecture	88.75%	0.925	0.940	0.932
SVM (Augmented)	RBF kernel	83.50%	0.840	0.840	0.840
SVM (Non-augmented)	RBF kernel	83.13%	0.825	0.825	0.824
Random Forest (Non-augmented)	50 estimators	81.13%	0.841	0.841	0.841
Random Forest (Augmented)	100 estimators	80.25%	0.817	0.817	0.817
KNN (Non-augmented)	k=5	70.75%	0.729	0.731	0.727
KNN (Augmented)	k=9	69.37%	0.730	0.730	0.727

Best Model Architecture (CNN)

Input (64×64×1)
  ↓
Conv2D (32 filters, 3×3) + ReLU
  ↓
MaxPooling2D (2×2)
  ↓
Conv2D (64 filters, 3×3) + ReLU
  ↓
MaxPooling2D (2×2)
  ↓
Conv2D (128 filters, 3×3) + ReLU
  ↓
MaxPooling2D (2×2)
  ↓
Flatten
  ↓
Dense (128 units) + ReLU
  ↓
Dense (1 unit) + Sigmoid

Project Structure

CS464_Project/
├── code.ipynb              # Main Jupyter notebook with all experiments
├── docs/
│   └── report.pdf          # Comprehensive project report (26 pages)
├── Car-Bike-Dataset/       # Dataset directory (not included in repo)
│   ├── Car/
│   └── Bike/
├── pyproject.toml
├── .gitignore
└── README.md

Methodology

1. Data Preprocessing

• Image resizing to 64×64 pixels
• Grayscale conversion
• Normalization
• Train-validation-test split (60-20-20)

2. Data Augmentation

Applied transformations:

• Horizontal flips (50% probability)
• Gaussian blur (σ ∈ [0, 0.5])
• Additive Gaussian noise
• Random order application

3. Hyperparameter Tuning

KNN: Tested k values from 5 to 37 (step=2)

• Best non-augmented: k=5
• Best augmented: k=9

Random Forest: Tested estimators [10, 20, 50, 100, 150]

• Best both: n_estimators=50

SVM: Tested kernels [linear, poly, rbf, sigmoid]

• Best both: RBF kernel

CNN: Tested 4 different architectures

• Best: 4-layer architecture with additional dense layer

4. PCA Analysis

Tested component counts: [1, 50, 100, 250, 500, 1000, 2400]

• Optimal for KNN/SVM: 50-100 components
• Found diminishing returns beyond 100 components