🛣️ Road Damage Classification Using ResNet50, InceptionV3, and VGG16

A Deep Learning Approach for Multi-Label Road Damage Detection

This project focuses on developing a robust multi-label classification system to identify and categorize various types of road damage. By leveraging pre-trained deep learning models—ResNet50, InceptionV3, and VGG16—we aim to automate the process of road damage detection from a dataset of 4,000 TIF images, each labeled with 26 different damage types. The system predicts severity on a scale of 0 to 5, aiding in road maintenance planning and resource allocation.

---

🎯 Objectives

• Develop a Multi-Label Classification System
  Predict multiple road damage types simultaneously using transfer learning.

• Model Comparison
  Evaluate ResNet50, InceptionV3, and VGG16 based on accuracy, precision, recall, and F2-score.

• Insightful Visualizations
  Generate loss curves, confusion matrices, and training metrics to compare model performance.

---

🗂️ Dataset

• Images: 4,000 .TIF images of roads
• Labels: 26 damage categories per image, each rated from 0 (none) to 5 (severe)

📩 For access to the dataset, please email: 25100147@lums.edu.pk

---

🧠 Methodology

🔍 Data Preprocessing

• Read and process multi-label .TIF road images
• Resize and normalize input images
• Prepare label tensors for 26-class severity scores

🤖 Model Architectures

Model	Highlights
ResNet50	Deep residual architecture helps mitigate vanishing gradients
InceptionV3	Auxiliary classifiers improve gradient flow during training
VGG16	Simpler and deeper, but more resource-intensive

---

🏋️‍♂️ Training & Evaluation

• Loss Function: CrossEntropyLoss
• Optimizer: Adam
• Validation after each epoch
• Metrics Tracked: Accuracy, Precision, Recall, F2-score

📊 Visualizations

• 📉 Loss Curves for training vs. validation
• 📋 Confusion Matrices for detailed error analysis
• 📈 Metric Trends across epochs for all models

---

🧪 Hyperparameter Tuning

• Tuned:
  • Learning Rates
  • Batch Sizes
  • Layer Freezing/Unfreezing
• Goal: Maximize validation F2-score

---

🚀 Deployment Potential

• Application of best model in:
  • Automated road inspection systems
  • Government road maintenance scheduling
  • Scalable infrastructure monitoring

---

🏁 Results Summary

Model	Strengths	Notes
ResNet50	Balanced across all metrics	Efficient training and decent generalization
InceptionV3	Best Precision & F2-score	Benefit from auxiliary classifiers
VGG16	High Recall	Computationally heavier than others

---

📌 Conclusions

• InceptionV3 shows the best performance in precision and F2-score.
• All models demonstrate the viability of deep learning for road damage detection.
• This approach is highly scalable and offers real-world impact in smart city initiatives.

---

🔮 Future Work

• 🔓 Unfreeze More Layers to fine-tune deeper representations
• 📈 Expand Dataset for better generalization
• 🌐 Deploy model to mobile/web for real-time road damage detection

---

👨‍💻 Author & Contact

Developed by Muhammad Saad Haroon
For dataset or inquiries, contact: 25100147@lums.edu.pk

---

Built with PyTorch and passion for practical AI.