🏗️ Structural-Defects-Network-MLOps

An end-to-end MLOps pipeline for detecting concrete structural cracks using deep learning (ResNet and MobileNet) with PyTorch, Airflow for orchestration, and MLflow for experiment tracking. In addition, we support

Problem Statment

As the recent earthquake in Thailand caused significant structural damage, particularly to high-rise buildings which led to widespread concrete cracking and breakage. Therefore, the need for efficient and reliable damage assessment has become more critical than ever. Structural defects such as cracks and spalling not only undermine the integrity of infrastructure but also pose serious risks to public safety if not detected and repaired promptly

This project proposes an AI-powered defect detection system that leverages deep learning to automate the identification and localization of structural damage from images. By utilizing pre-trained models such as ResNet, EfficientNet and MobileNetV3. The system classifies images as either cracked or non-cracked and applies object detection techniques for precise localization. This approach not only improves the speed and accuracy of structural assessments but also supports engineers and decision-makers in prioritizing repairs, ultimately enhancing infrastructure resilience after natural disasters.

In order to implement this project, we use the dataset from Kaggle (URL: https://www.kaggle.com/datasets/aniruddhsharmastructural-defects-network-concrete-crack-images). This dataset is called "Structural Defects Network (SDNET) 2018".

---

🔧 Prerequisites

• Python >= 3.10
• Docker + Docker Compose
• Optional: Virtualenv or conda

---

✨ System Design

📦 Setup Instructions

1. Create a Virtual Environment (Recommended)

Windows

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

macOS/Linux

python -m venv env
source env/bin/activate

2. Install Dependencies

pip install -r requirements.txt

---

3. Run Preprocessing

Run the preprocessing script before training or evaluation:

python src/preprocess_set_seed.py

4. Data Artifacts

After running the preprocessing script: The train, test, and validation datasets will be saved inside the artifact_folder/ directory. The artifact_folder/ is excluded from version control and will not be uploaded to Git (as specified in .gitignore).

---

🏢 Project Structure

Structural-Defects-Network-MLOps/
├── Dataset/                  # Raw Data (Decks, Walls, Pavements)
├── airflow/
│   ├── dags/                 # Airflow DAGs and Python callables
│   │   ├── main_dag.py
│   │   └── callables/        # All python callables function in main dag
│   ├── config/               # airflow.cfg and other configs
│   ├── plugins/              # Airflow plugins (if needed)
|   ├── mlflow/               # MLflow artifacts folder (will be created after run)
|   ├── model/                # local shared volume for saving models
│   ├── logs/                 # Airflow logs
│   ├── Dockerfile            # Custom Image for this project
│   └── docker-compose.yaml   # Docker Compose for Airflow stack
├── artifact_folder/          # Preprocessed datasets (train/val/test) – auto-generated after preprocessing
├── src/                      # CLI scripts for local training or upload
│   ├── scripts/              # All extra python scripts
│   ├── preprocess_set_seed.py # preprocessing scripts to run before using jupyter notebooks
│   ├── MobileNetV3.ipynb     # MobileNetv3 training steps
│   ├── resnet.ipynb          # ResNet18 training steps
│   └── docker-compose.yaml   # Docker Compose for Airflow stack
├── model/                    # Saved PyTorch model files                  # Airflow and pipeline logs
├── requirements.txt
└── README.md

---

🧹 Data Preprocessing

Run this step before training:

python src/preprocess_set_seed.py

This will generate train/, val/, and test/ image folders + CSVs under artifact_folder/.

---

🐳 Run with Docker + Airflow + MLFlow

1. Add Environment Variables

Create a .env in the airflow folder root airflow/:

AIRFLOW_UID=50000
CLOUDINARY_CLOUD_NAME=your_cloud_name
CLOUDINARY_API_KEY=your_key
CLOUDINARY_API_SECRET=your_secret
DATASET_PATH=<path>/<to>/your/projectStructural-Defects-Network-MLOps\Dataset
MLFLOW_TRACKING_URI=http://mlflow:5000

| Reminder: this dataset path should be absolute path.

2. Start Airflow + MLflow

cd airflow
docker compose up -d --build

Airflow UI → http://localhost:8080 MLFlow UI -> http://localhost:5000

Username / Password: airflow / airflow

With this airflow and mlflow, you can run whole our Structural Defect Machine Learning Pipeline in our airflow and monitor model performance and parameters in MLFlow instantly.

3. Stop Airflow

docker compose down --volumes --rmi all

---

Logs from training tasks in Airflow will be registered automatically (ResNet, MobileNet).

---

⚙️ Triggering the DAG

You can trigger the DAG manually from the Airflow UI, or let it run based on the monthly schedule (@weekly).

It checks:

• If new images have been uploaded to Cloudinary (e.g. under Users/Cracked/, Users/Non-cracked/)
• If >=10 new images are found → preprocess → train → log to MLflow
• Otherwise → skip training

---

✨ Notes

• The dataset must follow this structure:

  Dataset/
  ├── Decks/
  ├── Pavements/
  └── Walls/
  

• Preprocessing includes augmentation, resizing, and denoising filters.

• Models are saved to /opt/airflow/model/ and registered via MLflow.

• Metrics are saved using XCom and optionally pushed to Airflow Variables.

---

🖼️ Frontend Prediction Interface

This project includes a simple React-based frontend web app that enables:

• 🔍 Uploading images for real-time prediction (Cracked / Non-cracked)
• 🧠 Model inference using the latest trained model (ResNet or MobileNet)
• 📤 Submitting images with labels for continuous learning

---

📁 Frontend Directory Structure

app/
├── templates/
├── demo_run.ipynb                      # Contains REACT_APP_API_URL
├── Dockerfile
├── main.py
└── requirements.txt

---

🚀 Quick Start

🐳 By Docker

1. Navigate to the frontend directory:

cd app

2. Build the docker image

docker build . --tag 'structure_prediction_ui'

3. Run the docker image and access the web

docker run -p 8000:8000 structure_prediction_ui

or use docker compose up --build

💨 By fastapi cli

1. Setup the virtual environment

python -m venv .venv
source .venv/bin/activate

2. Install the dependencies

pip install -r requirements.txt

3. Access the website:

mkdir static
fastapi dev main.py
localhost:8000

---

🌐 Features

• Drag-and-Drop Upload or file selector

• Calls backend API (/predict) to return:

  • Predicted label
  • Confidence score

• Option to submit the image to Cloudinary under:

---

• Users/Cracked/

• Users/Non-cracked/

• Those images will later be retrained during the next Airflow pipeline run

---

🧠 How It Works

• The frontend sends an image to a FastAPI

• The backend:

  • Loads the latest model checkpoint
  • Applies preprocessing
  • Performs a forward pass
  • Returns prediction + confidence

• Optionally, the backend can upload the labeled image to Cloudinary for future training

---

🧪 API Routes (Backend)

Method	Endpoint	Description
GET	/	Predicts the class of a user image
POST	/	Predict the class of a user image

---

How to launch the UI?

🐳 By Docker

1. Navigate to the frontend directory:

cd app

2. Build the docker image

docker build . --tag 'structure_prediction_ui'

3. Run the docker image and access the web

docker run -p 8000:8000 structure_prediction_ui

Dry run

Access the website:

mkdir static
fastapi dev main.py
localhost:8000

---

📝 TODOs (Optional)

• Add a history view of past predictions
• Display class probabilities in a bar chart
• Show model version info

---

📄 License

MIT License