structural-damage-detection

🧱 Early Detection of Structural Damage Using Deep Learning & Computer Vision

📌 Project Overview

This project focuses on early detection of structural damage (cracks, spalling, collapse risk) in walls and concrete structures using a combination of Deep Learning (MobileNet CNN) and Computer Vision (OpenCV + image processing) techniques.

The system provides:

• Crack / Non‑crack classification
• Damage severity analysis
• Crack visualization (minor & major cracks)
• Training performance metrics (loss & accuracy)
• Interactive Streamlit web application with upload & camera support

This project is suitable for academic projects, research prototypes, and real‑world inspection assistance systems.

---

🎯 Key Objectives

• Detect cracks and damage in wall/structure images
• Ignore irrelevant objects (people, windows, background clutter)
• Highlight minor cracks (green) and major cracks (red)
• Provide interpretable metrics (edge density, texture variance, confidence)
• Visualize model training performance

---

🧠 Why This Approach?

Why CNN (MobileNet)?

• Lightweight and fast
• Works well on limited hardware
• Pretrained on ImageNet → strong feature extraction
• Suitable for real‑time and deployment

Why OpenCV + Image Processing?

• Crack geometry is thin and irregular → classical CV works well
• Edge detection & skeletonization help localize cracks
• Enhances explainability beyond black‑box ML

Why Hybrid (ML + CV)?

• CNN → classification (cracked / non‑cracked)
• OpenCV → localization & severity analysis
• Produces both decision + visualization

---

🗂️ Datasets Used

1️⃣ Crack500 Dataset

Used for training the CNN model.

• Contains cropped crack images
• High‑quality labeled crack patterns

2️⃣ SDNET2018 Dataset

Used for realistic structure damage analysis.

• Categories:

  • Walls
  • Pavements
  • Decks

• Each with:

  • Cracked
  • Non‑cracked

📁 Dataset Structure (Required)

datasets/
├── train/
│   ├── Cracked/
│   └── Non-cracked/
├── val/
│   ├── Cracked/
│   └── Non-cracked/

⚠️ Only this structure works with ImageFolder in PyTorch.

---

🤖 Models Used

🔹 MobileNetV2 (PyTorch)

• Pretrained weights: ImageNet
• Modified classifier layer for binary classification
• Saved model: mobilenet_crack.pth

🔹 Grad‑CAM

• Used for visual explanation of CNN focus regions

---

🔬 Training Process

Steps:

1. Load dataset using ImageFolder
2. Apply data augmentation
3. Load pretrained MobileNetV2
4. Replace classifier head
5. Train using CrossEntropyLoss
6. Track loss & accuracy per epoch

7. Save:

   • Model weights
   • Training loss graph
   • Training accuracy graph

Output:

outputs/
├── loss.png
├── accuracy.png

---

🖼️ Image Analysis Process (Inference)

1. Image upload / camera capture
2. Object filtering (ignore people/windows)

3. CNN prediction:

   • Cracked / Non‑cracked
   • Confidence score

4. OpenCV analysis:

   • Edge density
   • Texture variance (GLCM)
   • Damage severity

5. Crack highlighting:

   • Minor cracks → Green
   • Major crack → Red (longest skeleton path)

---

🌐 Streamlit Web Application

Features

• Upload image
• Camera capture
• Training metrics display (front page)
• Real‑time crack analysis
• Highlighted crack visualization

Main File

app.py

---

▶️ How to Run the Project

1️⃣ Clone / Open Project

cd "project c"

2️⃣ Create Virtual Environment

python -m venv venv
.\venv\Scripts\activate

3️⃣ Install Dependencies

pip install -r requirements.txt

4️⃣ Train Model (Optional)

python train.py

5️⃣ Run Web App

streamlit run app.py

---

🧾 Software Requirements

• Windows / Linux
• Python 3.9+
• VS Code (recommended)
• Streamlit

Libraries Used

• torch, torchvision
• opencv‑python
• numpy
• matplotlib
• scikit‑image
• streamlit
• pillow

---

💻 Hardware Requirements

• Minimum:

  • CPU i5 / Ryzen 5
  • 8 GB RAM

• Recommended:

  • GPU (CUDA supported)
  • 16 GB RAM

---

🌲 Project Structure

project c/
├── app.py
├── train.py
├── evaluate.py
├── requirements.txt
├── README.md
├── datasets/
│   ├── crack500/
│   └── sdnet/
├── models/
│   ├── mobilenet_model.py
│   └── gradcam.py
├── utils/
│   ├── classifier.py
│   ├── cv_analysis.py
│   ├── image_utils.py
│   └── object_filter.py
├── outputs/
│   ├── loss.png
│   └── accuracy.png
└── venv/

---

⚠️ Things to Keep in Mind

✅ Do

• Activate venv before running
• Maintain dataset folder structure
• Run train.py before first inference
• Use clear wall/structure images

❌ Don’t

• Rename dataset folders randomly
• Mix cracked & non‑cracked images
• Delete outputs/ after training
• Upload unrelated images (people, objects)

---

🚀 Future Improvements

• Add YOLO‑based object filtering
• Multi‑class damage classification
• Crack width & length measurement
• Video / live camera stream analysis
• Cloud deployment
• Automated report generation (PDF)
• Improved major crack confidence scoring

---

🏁 Conclusion

This project demonstrates a practical, explainable, and deployable approach for structural damage detection using modern AI and classical vision techniques. It balances accuracy, performance, and interpretability, making it suitable for real‑world inspection systems.

---

📌 Developed as an academic and research‑oriented structural inspection system.