AI-Based Anomaly Detection in Industrial Screws

---

🧠 Project Overview

This project aims to automate visual defect detection in industrial screw images using a deep learning autoencoder. The system reconstructs input images and flags anomalies based on reconstruction error — offering a faster, more consistent alternative to manual inspection.

---

🎓 Academic Context

This work was developed as part of the Deep Learning Project at Esprit School of Engineering, under the supervision of Professor Hannen Ennine.

---

✨ Features

✅ Web Interface

• Upload image for real-time defect classification
• Dynamic prediction result with confidence score
• Clean UI built with Tailwind CSS

✅ Backend

• FastAPI handles image upload and model inference
• TensorFlow Autoencoder detects anomalies based on reconstruction error
• Confidence score based on how far the reconstruction error is from a learned threshold

✅ Model

• CNN-based autoencoder
• Trained only on “good” samples
• Anomalies are detected based on deviation from learned patterns

---

⚙️ Setup

# Clone the repo
git clone https://github.com/RideneFiras/anomaly_detection.git
cd anomaly_detection

# Create and activate a virtual environment
python3.11 -m venv venv
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Run the app
make run

Then open your browser at http://127.0.0.1:8000

---

🗂️ Project Structure

├── app.py                  # FastAPI backend
├── index.html              # Frontend web UI (served via FastAPI)
├── autoencoder_model.h5    # Trained autoencoder model
├── screw/                  # MVTec dataset folder (train/test)
│   └── screw/train/good/   # Used during training
│   └── screw/test/         # Used for evaluation
├── assets/                 # Contains demo.gif and other visuals
├── requirements.txt        # Python dependencies
├── README.md               # You're here!

Download the screw dataset from Mvtec AD dataset

🧪 Model Logic

• Input image is resized and normalized
• Autoencoder reconstructs the image
• Compute mean squared error (MSE) between original and reconstruction
• If error > predefined threshold → classify as Defective
• Otherwise → Good

---

🧠 Training Details

• Trained on only good images from the MVTec screw dataset
• Validation split: 20%
• Loss: Mean Squared Error (mse)
• Optimizer: Adam (lr=0.0001)
• Regularization: Dropout, L2

---

🧾 Results

• ✅ Accuracy: 99%
• ✅ Precision/Recall: 98%–100%
• ✅ Threshold: ~0.0148 MSE
• ✅ Real-time prediction via FastAPI

---

✅ Technologies Used

• Python
• TensorFlow / Keras
• FastAPI
• OpenCV
• TailwindCSS
• HTML & JavaScript

---

📜 License

This project is for academic purposes only and was developed as part of coursework at Esprit School of Engineering.