GitHub - dinraj910/Structural-Damage-Severity-Assessment-CNN: A production-ready web application that deploys a Convolutional Neural Network for automated visual assessment of structural damage in building photographs. This AI-assisted tool can screen thousands of structures in minutes, guiding human inspectors to the most critical cases first.

Overview • Features • Architecture • Quick Start • Demo • Tech Stack • Performance • Roadmap • Contributing • License

🎯 Overview

💡 What It Does

A production-ready Computer Vision system that classifies building and infrastructure images into four structural damage severity levels using a fine-tuned ResNet50 CNN — deployed as an interactive Streamlit web application with Grad-CAM explainability.

🌍 Why It Matters

After natural disasters — earthquakes, hurricanes, floods — rapid structural triage saves lives. Traditional manual assessment requires expert engineers and days of fieldwork. This AI system can screen thousands of structures in minutes, prioritising critical cases for human inspectors and accelerating disaster response.

🏷️ Damage Severity Classes

Class	Severity	Icon	Description	Use Case
`0`	No Damage	✅	Structure appears fully intact	Clear for occupancy
`1`	Minor Damage	⚠️	Surface-level cracks, cosmetic issues	Monitor & schedule repair
`2`	Moderate Damage	🔶	Visible structural compromise	Restrict access, urgent repair
`3`	Severe Damage	🔴	Major structural failure or collapse	Evacuate, demolition review

✨ Features

Feature	Description	Status
📤 Image Upload	Drag & drop JPG/PNG building photos	✅
🤖 Real-Time Inference	Instant 4-class severity prediction	✅
📊 Confidence Scores	Per-class softmax probability breakdown	✅
🔍 Grad-CAM Explainability	Visual heatmap of decision-driving regions	✅
🎨 Professional Dark UI	Custom CSS metric cards & probability bars	✅
⚡ Cached Model Loading	`@st.cache_resource` — zero reload latency	✅
🧩 Modular Codebase	Clean separation: UI / preprocessing / inference / XAI	✅
📘 Interpretation Guide	Built-in confidence thresholds & disclaimers	✅
☁️ Cloud-Ready	One-click deploy to Streamlit Cloud	✅
🎛️ Interactive Controls	Adjustable Grad-CAM opacity slider	✅

🏗️ Architecture

🧠 Model Architecture — ResNet50 Transfer Learning

                    ┌─────────────────────────────────────────────┐
                    │              INPUT IMAGE                     │
                    │           224 × 224 × 3 (RGB)               │
                    └─────────────────┬───────────────────────────┘
                                      │
                                      ▼
                    ┌─────────────────────────────────────────────┐
                    │         RESNET50 BACKBONE                    │
                    │      (ImageNet Pre-trained, Frozen)          │
                    │                                              │
                    │   Conv1 ──► Conv2 ──► Conv3 ──► Conv4 ──►   │
                    │                                              │
                    │   ──► Conv5 (conv5_block3_out)  ◄── Grad-CAM│
                    └─────────────────┬───────────────────────────┘
                                      │
                                      ▼
                    ┌─────────────────────────────────────────────┐
                    │       GLOBAL AVERAGE POOLING 2D              │
                    │            (2048,)                           │
                    └─────────────────┬───────────────────────────┘
                                      │
                                      ▼
                    ┌─────────────────────────────────────────────┐
                    │          DENSE (256, ReLU)                    │
                    └─────────────────┬───────────────────────────┘
                                      │
                                      ▼
                    ┌─────────────────────────────────────────────┐
                    │          DROPOUT (0.5)                        │
                    └─────────────────┬───────────────────────────┘
                                      │
                                      ▼
                    ┌─────────────────────────────────────────────┐
                    │     DENSE (4, Softmax) ──► OUTPUT             │
                    │                                              │
                    │   [No Damage, Minor, Moderate, Severe]       │
                    └─────────────────────────────────────────────┘

⚙️ Application Pipeline — End-to-End Flow

  ┌──────────────┐     ┌──────────────────┐     ┌──────────────────┐
  │              │     │                  │     │                  │
  │   USER       │────►│   STREAMLIT      │────►│  PREPROCESSING   │
  │  UPLOADS     │     │   FRONTEND       │     │                  │
  │  IMAGE       │     │   (app.py)       │     │  • Decode bytes  │
  │              │     │                  │     │  • Resize 224²   │
  └──────────────┘     └──────────────────┘     │  • Normalize /255│
                                                 │  • Batch expand  │
                                                 └────────┬─────────┘
                                                          │
                                                          ▼
  ┌──────────────┐     ┌──────────────────┐     ┌──────────────────┐
  │              │     │                  │     │                  │
  │   RESULTS    │◄────│   GRAD-CAM       │◄────│   CNN MODEL      │
  │   DISPLAY    │     │   EXPLAINABILITY  │     │   INFERENCE      │
  │              │     │                  │     │                  │
  │  • Class     │     │  • Heatmap gen   │     │  • Forward pass  │
  │  • Score     │     │  • Overlay blend │     │  • Softmax probs │
  │  • Heatmap   │     │  • Visualization │     │  • argmax class  │
  │              │     │                  │     │                  │
  └──────────────┘     └──────────────────┘     └──────────────────┘

🔬 Technical Deep Dive

📐 Preprocessing Pipeline

The image preprocessing mirrors the exact pipeline used during model training:

# 1. Decode uploaded bytes → PIL RGB Image
image = Image.open(BytesIO(raw_bytes)).convert("RGB")

# 2. Resize to model input dimensions
image_resized = image.resize((224, 224), Image.LANCZOS)

# 3. Convert to float32 array & normalize to [0, 1]
img_array = np.asarray(image_resized, dtype=np.float32) / 255.0

# 4. Add batch dimension → (1, 224, 224, 3)
img_array = np.expand_dims(img_array, axis=0)

🔥 Grad-CAM Implementation

Grad-CAM (Selvaraju et al., ICCV 2017) produces visual explanations by:

Extracting activations from conv5_block3_out (the last ResNet50 conv block)
Computing gradients of the predicted class w.r.t. these activations
Global-average-pooling gradients to obtain per-filter importance weights
Combining weighted activations into a spatial heatmap
Overlaying the heatmap on the original image with adjustable opacity

grad_model = tf.keras.models.Model(
    inputs=model.inputs,
    outputs=[model.get_layer("conv5_block3_out").output, model.output]
)

with tf.GradientTape() as tape:
    conv_outputs, predictions = grad_model(img_array)
    class_channel = predictions[:, pred_index]

grads = tape.gradient(class_channel, conv_outputs)
pooled_grads = tf.reduce_mean(grads, axis=(0, 1, 2))

heatmap = conv_outputs[0] @ pooled_grads[..., tf.newaxis]
heatmap = tf.maximum(heatmap, 0) / tf.math.reduce_max(heatmap)

⚡ Model Caching Strategy

The model is loaded using Streamlit's @st.cache_resource decorator, which:

Loads the .keras model file once at startup
Shares the same model instance across all user sessions
Eliminates redundant disk I/O and GPU memory allocations
Persists across Streamlit re-runs (script re-execution)

@st.cache_resource(show_spinner="Loading CNN model …")
def load_model(model_path=None):
    model = tf.keras.models.load_model(model_path)
    return model

📁 Project Structure

Structural-Damage-Severity-Assessment-CNN/
│
├── 🎯 app.py                          # Streamlit UI — entry point
├── 📋 requirements.txt                # Pinned Python dependencies
├── 📖 README.md                       # This file
├── 🚫 .gitignore                      # Git ignore rules
│
├── 🤖 model/
│   └── structural_damage_cnn.keras    # Trained ResNet50 weights (~100 MB)
│
├── 🧰 utils/
│   ├── __init__.py                    # Package exports
│   ├── preprocessing.py               # Image decoding, resize, normalize
│   ├── prediction.py                  # Cached model loader & inference
│   └── gradcam.py                     # Grad-CAM heatmap & overlay
│
├── 🖼️ assets/
│   ├── icons/                         # App icons & branding
│   └── sample_images/                 # Example test images
│
├── 📓 notebooks/
│   ├── Structural_Damage_Assessment.ipynb  # Full training notebook
│   └── structural_damage_assessment.py     # Notebook as script
│
├── 📸 screenshots/                    # App screenshots for docs
│
└── 🎨 .streamlit/
    └── config.toml                    # Dark theme & server config

🚀 Quick Start

📋 Prerequisites

Requirement	Version	Purpose
Python	3.10+	Runtime
pip	Latest	Package management
Git	Any	Repository cloning
GPU (optional)	CUDA 11.8+	Accelerated inference

📥 Installation

# ── 1. Clone the repository ──────────────────────────────────────────────────
git clone https://github.com/dinraj910/Structural-Damage-Severity-Assessment-CNN.git
cd Structural-Damage-Severity-Assessment-CNN

# ── 2. Create & activate virtual environment ─────────────────────────────────
python -m venv venv

# Linux / macOS
source venv/bin/activate

# Windows
venv\Scripts\activate

# ── 3. Install dependencies ──────────────────────────────────────────────────
pip install -r requirements.txt

▶️ Run the Application

streamlit run app.py

The app launches at http://localhost:8501 — upload a building image and get instant results.

🖥️ Screenshots

� Application Interface

🏠 Landing Page
Clean, informative interface explaining the model, classes, and architecture

📤 Upload Interface
Drag & drop image upload with sidebar controls

🔮 Prediction Results

📊 Classification Results
Severity prediction with confidence scores and probability breakdown

📈 Confidence Breakdown
Per-class probability bars with color-coded severity levels

🔍 Grad-CAM Explainability

🎯 Activation Heatmap
Raw Grad-CAM heatmap showing decision-driving regions

🌈 Grad-CAM Overlay
Heatmap overlaid on original image with adjustable opacity

🔬 Advanced Features

⚙️ Interactive Controls
Sidebar options and explainability settings

📋 Complete Analysis
Full end-to-end damage assessment workflow

⚙️ Configuration

Streamlit Theme (`.streamlit/config.toml`)

Setting	Value	Description
`primaryColor`	`#4fc3f7`	Accent colour (light blue)
`backgroundColor`	`#0e1117`	App background (dark)
`secondaryBackgroundColor`	`#1a1a2e`	Sidebar & card background
`textColor`	`#e0e0e0`	Primary text colour
`font`	`sans serif`	Typography
`maxUploadSize`	`10 MB`	Maximum upload file size

Environment Variables

Variable	Required	Default	Description
`MODEL_PATH`	❌	`model/structural_damage_cnn.keras`	Custom model file path

🛠️ Tech Stack

Layer	Technology	Role
	Python 3.10+	Core language
	TensorFlow 2.15	Deep learning framework
	Keras	High-level model API
	OpenCV	Image processing & Grad-CAM overlay
	NumPy	Numerical computation
	Streamlit	Web application framework
🖼️	Pillow	Image decoding & manipulation
📊	Matplotlib	Visualisation support

🧠 Skills Demonstrated

Skill Area	Techniques Applied
Deep Learning	Transfer Learning, Fine-Tuning, CNNs
Computer Vision	Image Classification, Feature Extraction
Explainable AI	Grad-CAM, Gradient Visualization
Model Deployment	TensorFlow SavedModel, Keras `.keras` format
Web Development	Streamlit, Custom CSS, Responsive Design
Software Engineering	Modular Architecture, Caching, Type Hints
MLOps	Model Versioning, Cloud Deployment, `.gitignore`

📈 Performance

Metric	Value	Notes
Training Accuracy	~93%	10 epochs, ResNet50 frozen backbone
Validation Accuracy	~89%	20% hold-out split
Inference Time	~120ms	CPU (single image, cached model)
Model Size	~100 MB	ResNet50 + custom head (`.keras`)
Input Resolution	224 × 224	RGB, normalised to [0, 1]
Output Classes	4	Softmax probability distribution

📝 Metrics are approximate and depend on hardware. GPU inference is significantly faster.

🗺️ Roadmap

 ✅ v1.0          🔄 v1.1          🔄 v1.2          🔄 v1.3          🔄 v2.0          🔜 v2.1          🔜 v3.0
┌─────────┐    ┌─────────┐    ┌─────────┐    ┌─────────┐    ┌─────────┐    ┌─────────┐    ┌─────────┐
│  Core   │───►│  Batch  │───►│   PDF   │───►│  Multi  │───►│   API   │───►│ Mobile  │───►│  Edge   │
│   App   │    │ Upload  │    │ Report  │    │  Model  │    │ Docker  │    │Responsive│   │ Deploy  │
└─────────┘    └─────────┘    └─────────┘    └─────────┘    └─────────┘    └─────────┘    └─────────┘
  🟢 DONE        🔵 PLANNED      🔵 PLANNED      🔵 PLANNED      🔵 PLANNED      🟣 FUTURE       🟣 FUTURE

Version	Feature	Status
v1.0	Core App — Upload, Predict, Grad-CAM	✅ Done
v1.1	Batch image upload & CSV export	🔄 Planned
v1.2	Automated PDF damage report generation	🔄 Planned
v1.3	Multi-model ensemble (ResNet + EfficientNet)	🔄 Planned
v2.0	REST API + Docker containerisation	🔄 Planned
v2.1	Mobile-responsive layout	🔜 Future
v3.0	Edge deployment (TFLite / ONNX)	🔜 Future

☁️ Cloud Deployment

🚀 Deploy to Streamlit Cloud (Free)

Push this repository to GitHub
Visit share.streamlit.io
Connect your GitHub repo
Select app.py as the entry point
Deploy — dependencies install automatically from requirements.txt

⚠️ Ensure model/structural_damage_cnn.keras is committed (or use Git LFS for large files).

🐳 Docker Deployment

FROM python:3.10-slim

WORKDIR /app
COPY . .

RUN pip install --no-cache-dir -r requirements.txt

EXPOSE 8501

CMD ["streamlit", "run", "app.py", "--server.port=8501", "--server.address=0.0.0.0"]

docker build -t structural-damage-assessment .
docker run -p 8501:8501 structural-damage-assessment

🤝 Contributing

Contributions are welcome! Here's how:

# 1. Fork the repository
# 2. Create a feature branch
git checkout -b feature/amazing-feature

# 3. Commit your changes
git commit -m "feat: add amazing feature"

# 4. Push to your branch
git push origin feature/amazing-feature

# 5. Open a Pull Request

📜 Contribution Guidelines

Follow existing code style and modular architecture
Add docstrings to all new functions
Test changes locally before submitting
Use conventional commit messages (feat:, fix:, docs:, etc.)

⚠️ Disclaimer

This application is a decision-support tool and does NOT replace professional structural engineering assessments. Predictions are generated by a machine learning model and may be incorrect — particularly for building types, lighting conditions, or architectural styles not represented in the training data. Do not use this system as the sole basis for safety-critical decisions. Always consult a qualified structural engineer.

📄 License

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

MIT License — Copyright (c) 2026

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software.

👤 Author

DINRAJ K DINESH

Deep Learning Engineer · Computer Vision Researcher

🙏 Acknowledgements

Resource	Citation
ResNet50	He et al., "Deep Residual Learning for Image Recognition" — CVPR 2016
Grad-CAM	Selvaraju et al., "Grad-CAM: Visual Explanations from Deep Networks" — ICCV 2017
Concrete Crack Dataset	Özgenel, Ç.F. — Concrete Crack Images for Classification
Hurricane Damage Dataset	Satellite Images of Hurricane Damage — Kaggle
Streamlit	Open-source framework for ML application deployment
TensorFlow	Google Brain — End-to-end ML platform

⭐ Star History

💖 Show Your Support

If this project helped you, please consider giving it a ⭐!

⭐ Star this repo · 🍴 Fork it · 🐛 Report bugs · 💡 Request features

_{Built with ❤️ and TensorFlow — Making structures safer, one prediction at a time.}

Name		Name	Last commit message	Last commit date
Latest commit History 3 Commits
.streamlit		.streamlit
assets/sample_images		assets/sample_images
notebooks		notebooks
screenshots		screenshots
utils		utils
.gitignore		.gitignore
README.md		README.md
app.py		app.py
requirements.txt		requirements.txt

Folders and files

Latest commit

History

Repository files navigation

🎯 Overview

💡 What It Does

🌍 Why It Matters

🏷️ Damage Severity Classes

✨ Features

🏗️ Architecture

🔬 Technical Deep Dive

📁 Project Structure

🚀 Quick Start

📋 Prerequisites

📥 Installation

▶️ Run the Application

🖥️ Screenshots

� Application Interface

🔮 Prediction Results

🔍 Grad-CAM Explainability

🔬 Advanced Features

⚙️ Configuration

Streamlit Theme (.streamlit/config.toml)

Environment Variables

🛠️ Tech Stack

🧠 Skills Demonstrated

📈 Performance

🗺️ Roadmap

☁️ Cloud Deployment

🤝 Contributing

📜 Contribution Guidelines

⚠️ Disclaimer

📄 License

👤 Author

🙏 Acknowledgements

⭐ Star History

💖 Show Your Support

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Streamlit Theme (`.streamlit/config.toml`)

Packages