Vehicle Damage Classification using Deep Learning

Overview

This project focuses on automated vehicle damage classification using deep learning. The model analyzes uploaded images of cars and predicts the type and location of damage, enabling applications in insurance automation, inspection systems, and fleet management.

The system is built as an end-to-end ML application, with DL methodological integrations of:

• Deep Learning model (ResNet50)
• FastAPI backend
• Streamlit frontend

---

Download Model

Due to GitHub file size limits, the trained model is hosted externally:

https://drive.google.com/drive/folders/1fhkW0NP0Dh_YLnjaaE0FGODOMhS7aIGN?usp=sharing

---

Problem Statement

Manual vehicle damage inspection is:

• Time-consuming
• Subjective
• Prone to human error

This project aims to:

Automate damage detection and classification using computer vision.

---

Approach & Methodology

1. Data Understanding

The dataset consists of vehicle images categorized into:

• Front Breakage
• Front Crushed
• Front Normal
• Rear Breakage
• Rear Crushed
• Rear Normal

This makes it a multi-class image classification problem (6 classes).

---

2. Data Preprocessing

Image Transformations

• Resize to 224×224 (standard input for ResNet)
• Normalize using ImageNet statistics

transforms.Normalize(
  mean=[0.485, 0.456, 0.406],
  std=[0.229, 0.224, 0.225]
)

💡 Why?

• Ensures compatibility with pretrained models
• Improves convergence and stability

---

3. Model Selection

🔹 ResNet50 (Transfer Learning)

We used a pretrained ResNet50 model and modified it:

• Froze initial layers → retain learned features
• Fine-tuned deeper layers → adapt to damage patterns
• Replaced final fully connected layer → match 6 classes

self.model = models.resnet50(weights='DEFAULT')

💡 Why ResNet50?

• Proven performance in image tasks
• Deep architecture with residual connections
• Prevents vanishing gradient problem

---

4. Training Strategy

🔁 Experiments Performed

• Baseline model (no tuning)
• Fine-tuning last layers
• Regularization using Dropout
• Hyperparameter tuning

⚙️ Optimization

• Loss: CrossEntropyLoss
• Optimizer: Adam / SGD (depending on experiment)

💡 Why Transfer Learning?

• Reduces training time
• Works well with limited data
• Leverages pretrained knowledge

---

5. Hyperparameter Tuning

Performed tuning to improve model performance:

• Learning rate
• Batch size
• Dropout rate

💡 Why?

To balance:

• Underfitting vs Overfitting
• Speed vs Accuracy

---

6. Model Output

The model predicts one of the following:

[
 'Front Breakage',
 'Front Crushed',
 'Front Normal',
 'Rear Breakage',
 'Rear Crushed',
 'Rear Normal'
]

---

🏗️ System Architecture

User Upload (Streamlit UI)
        ↓
Image Preprocessing
        ↓
Model Inference (ResNet50)
        ↓
Prediction Output
        ↓
Displayed on UI / API Response

---

Tech Stack

Component	Technology
DL Framework	PyTorch
Backend	FastAPI
Frontend	Streamlit
Image Handling	PIL
Deployment Ready	Render / Streamlit Cloud

---

Project Structure

project/
│
├── app.py               # Streamlit UI
├── server.py            # FastAPI backend
├── model_helper.py      # Model loading & prediction logic
├── saved_model.pth      # Trained model
├── requirements.txt
└── README.md

---

How to Run Locally

1. Clone Repository

git clone <your-repo-link>
cd project

2. Install Dependencies

pip install -r requirements.txt

3. Run Streamlit App

streamlit run app.py

4. Run FastAPI Server (optional)

uvicorn server:app --reload

---

API Endpoint

POST /predict

Input: Image file Output:

{
  "prediction": "Front Crushed"
}

---

Learnings

Through this project:

• Applied transfer learning in real-world scenario
• Built end-to-end ML system
• Integrated model with frontend & backend
• Understood deployment challenges

---

Conclusion

This project demonstrates how deep learning can be used to automate vehicle damage classification, reducing manual effort and improving efficiency in inspection systems.

---