🚗 Vehicle Damage Classification using Transfer Learning

A deep learning project that explores transfer learning for automatic vehicle damage classification using PyTorch.

Instead of training a convolutional neural network from scratch, this project fine-tunes multiple state-of-the-art pretrained models and compares their performance on a real-world multi-class vehicle damage dataset.

The goal of this repository is not only to achieve high accuracy but also to understand how different transfer learning architectures behave on the same dataset.

---

Project Overview

Vehicle damage assessment is one of the most common computer vision applications in the insurance and automobile industry. Manual inspection is time-consuming and subjective. This project investigates whether pretrained deep learning models can accurately classify different types of vehicle damage.

This repository currently includes experiments with:

• ResNet50
• DenseNet121
• EfficientNet-B0
• Vision Transformer (ViT-B/16)

More architectures will be added as the project progresses.

---

Dataset

• Total Images: 7,200
• Classes: 6

Damage Categories:

• Crack
• Scratch
• Tire Flat
• Dent
• Glass Shatter
• Lamp Broken

Dataset Split:

• Training: 80%
• Validation: 20%

---

Models Implemented

1. ResNet50

Transfer Learning

• ImageNet pretrained weights
• Layer4 unfrozen for fine-tuning
• Custom classification head
• PyTorch implementation

Results

Metric	Score
Accuracy	96.46%
Precision	96.48%
Recall	96.46%
F1 Score	96.45%

---

2. DenseNet121

Transfer Learning

• ImageNet pretrained weights
• Last Dense Block fine-tuned
• Custom classifier

Results

Metric	Score
Accuracy	96.32%
Precision	96.41%
Recall	96.32%
F1 Score	96.32%

---

3. EfficientNet-B0

Transfer Learning

• ImageNet pretrained weights
• Last feature block unfrozen
• Custom classifier

Results

Metric	Score
Accuracy	75.49%
Precision	74.02%
Recall	75.49%
F1 Score	74.72%

Although EfficientNet is known for excellent parameter efficiency, it underperformed on this dataset with the current fine-tuning strategy. This opens opportunities for future experiments involving data augmentation, longer training schedules, and hyperparameter tuning.

---

4. Vision Transformer (ViT-B/16)

Transfer Learning using the timm library.

• ImageNet pretrained weights
• Classification head fine-tuned
• Vision Transformer architecture

Results

Metric	Score
Accuracy	74.10%
Precision	74.48%
Recall	74.10%
F1 Score	73.62%

Although Vision Transformers achieve state-of-the-art results on very large datasets, they generally require more data and longer fine-tuning schedules than CNN-based models. On this relatively small dataset, ResNet50 and DenseNet121 significantly outperformed ViT.

---

Model Comparison

Model	Accuracy
🥇 ResNet50	96.46%
🥈 DenseNet121	96.32%
EfficientNet-B0	75.49%
Vision Transformer	74.10%

---

Tech Stack

• Python
• PyTorch
• Torchvision
• timm
• PIL
• NumPy
• Pandas
• scikit-learn

---

This project helped me understand much more than transfer learning.

Some of the concepts I explored while building these experiments include:

• PyTorch Dataset and DataLoader
• Transfer Learning
• Fine-Tuning
• Freezing and Unfreezing Layers
• CNN Feature Extraction
• Image Preprocessing
• Model Evaluation
• Precision, Recall and F1 Score
• Confusion Matrix
• Vision Transformers
• Comparative Model Analysis

---