🚗 Vehicle Type Image Classification

A deep learning project for classifying vehicle images into 5 categories using custom CNN architectures built with TensorFlow/Keras. Three model versions (v1, v2, v3) are trained and compared for performance.

---

📋 Table of Contents

• Overview
• Dataset
• Models
• Project Structure
• Requirements
• Usage
• Results
• License

---

Overview

This project builds and compares three custom Convolutional Neural Network (CNN) architectures to classify vehicle images. Each model is trained with data augmentation, early stopping, learning rate scheduling, and model checkpointing to ensure robust performance.

Classes:

• 🚗 Sedan
• 🚙 SUV
• 🛻 Pickup
• 🚘 Hatchback
• 🚌 Other

---

Dataset

The dataset is organized into class-named folders:

Vehicle Type Image Dataset/
├── Seden/
├── SUV/
├── Pickup/
├── Hatchback/
└── Other/

Split ratio:

• Training: 70%
• Validation: 15%
• Test: 15%

Preprocessing:

• Input image size: 224 × 224 × 3
• Pixel normalization: rescale = 1/255

Training Augmentation:

• Random rotation (±20°)
• Width & height shift (±20%)
• Horizontal flip
• Fill mode: nearest

---

Models

V1 — Lightweight CNN

A compact 3-block CNN suitable for fast training:

Layer	Filters	Notes
Conv2D	32	ReLU
MaxPool2D	—	2×2
Conv2D	64	ReLU
MaxPool2D	—	2×2
Conv2D	128	ReLU
MaxPool2D	—	2×2
Dense	128	ReLU + Dropout(0.5)
Dense	5	Softmax

---

V2 — Deep CNN with Batch Normalization

Deeper architecture with batch normalization and progressive dropout for stronger regularization:

• 4 conv blocks (32 → 64 → 128 → 256 filters), each with dual Conv2D + BatchNorm + MaxPool + Dropout
• Fully connected head: 512 → 256 → 5 (Softmax)

---

V3 — VGG-Style CNN

Inspired by VGGNet with stacked conv layers and larger dense layers:

• 4 conv blocks (64 → 128 → 256 → 512 filters), each with 2–3 Conv2D layers + MaxPool
• Fully connected head: 1024 → 512 → 5 (Softmax)

---

Training Configuration

All models share the same training setup:

Parameter	Value
Optimizer	Adam (lr=0.001)
Loss	Categorical Crossentropy
Epochs	50 (max)
Batch size	32
Early stopping	patience=10
LR reduction	factor=0.5, patience=5, min=1e-7
Checkpoint	Best val_accuracy saved

---

Project Structure

├── Vehicle Type Image Dataset/
│   ├── Hatchback/
│   ├── Other/
│   ├── Pickup/
│   ├── SUV/
│   └── Seden/
├── vehicle_type.ipynb          # Main notebook
├── v1_model_best.h5            # Saved best V1 model
├── v2_model_best.h5            # Saved best V2 model
├── v3_model_best.h5            # Saved best V3 model
├── v1_model_results.csv        # V1 training history
├── v2_model_results.csv        # V2 training history
├── v3_model_results.csv        # V3 training history
├── confusion_matrix_v1_model.png
├── confusion_matrix_v2_model.png
├── confusion_matrix_v3_model.png
└── README.md

---

Requirements

Python >= 3.8
tensorflow >= 2.x
numpy
pandas
matplotlib
seaborn
scikit-learn

Install all dependencies:

pip install tensorflow numpy pandas matplotlib seaborn scikit-learn

---

Usage

1. Clone the repository:

git clone https://github.com/Elsaraf1/vehicle-type-classification.git
cd vehicle-type-classification

2. Place the dataset in the root directory following the folder structure above.

3. Run the notebook:

jupyter notebook vehicle_type.ipynb

4. Load a saved model for inference:

import tensorflow as tf

model = tf.keras.models.load_model('v1_model_best.h5')
# or v2_model_best.h5 / v3_model_best.h5

predictions = model.predict(your_image_batch)

---

Results

Training curves and confusion matrices are saved automatically during training:

Model	Saved Weights	Training History	Confusion Matrix
V1	v1_model_best.h5	v1_model_results.csv	confusion_matrix_v1_model.png
V2	v2_model_best.h5	v2_model_results.csv	confusion_matrix_v2_model.png
V3	v3_model_best.h5	v3_model_results.csv	confusion_matrix_v3_model.png

---

License

This project is licensed under the MIT License. See the LICENSE file for details.