😷 Real-Time Face Mask Detection

A practical computer vision project that detects whether a person is wearing a mask or not in real time using PyTorch and OpenCV.
The system runs directly on laptop i7 CPU and includes a color-coded FPS counter to prove real-time performance.

---

Table of Contents

1. 🚀 Why This Matters
2. 🏗️ System Architecture
3. 📁 Project Structure
4. 📊 Performance Summary
5. 🔧 How It Works
6. 🖼️ Example Results
7. 🛠️ Requirements
8. 👤 Author

---

🚀 Why This Matters

Face mask detection is a widely used classification + deployment task in computer vision.
This project demonstrates:

• Real-time inference with PyTorch CNN (MobileNetV3-Small, transfer learning).
• Integration of OpenCV Haar cascades for face ROI extraction.
• Live FPS measurement to evaluate system viability on constrained hardware.

This project was completed during an 8-week internship at ARCH Technologies as part of the “Sharpening Your Hidden Skills for a Brighter Future” program.

---

🏗️ System Architecture

flowchart LR
    A[Webcam Input] --> B[Face Detection — Haar Cascade]
    B --> C[Crop & Preprocess Face ROI]
    C --> D[PyTorch CNN Model — MobileNetV3-Small]
    D --> E[Mask / No-Mask Prediction + Confidence]
    E --> F[Draw Label + Color-Coded FPS Counter]
    F --> G[Display Output Window]

Loading

---

📁 Project Structure

---

📊 Performance Summary

Metric	Value
Average FPS (Laptop CPU)	11.3 / 10.9
Resolution	640×480
Model	MobileNetV3-Small (transfer learning)
Latency per frame	~90 ms
Detection Accuracy	✅ Reliable on test dataset (masked vs unmasked faces)

• 🟢 >15 FPS = Real-time viable (Jetson / GPU)
• 🟡 10–15 FPS = Acceptable on CPU (this project’s range)
• 🔴 <10 FPS = Too slow

---

🔧 How It Works

👁️ Face Detection

• Uses OpenCV’s Haar Cascade to detect faces in each frame.
• Crops Regions of Interest (ROIs) for classification.

🧹 Preprocessing

• Resize face ROIs to 224×224.
• Normalize using ImageNet statistics (mean & std).

🧠 Model Training

• Transfer learning with MobileNetV3-Small in PyTorch.
• Fine-tuned last layers for binary classification: Mask vs No Mask.
• Metadata (normalization stats, labels) saved alongside model.

🚀 Real-Time Inference

• Predictions drawn as bounding boxes with class + confidence.
• FPS measured as 1 / (current_time - prev_time).
• FPS counter colors:
  • 🟢 >15 = Green (real-time ready)
  • 🟡 10–15 = Yellow (CPU-bound, acceptable)
  • 🔴 <10 = Red (not real-time)

---

🖼️ Example Results

✅ Detects mask vs no-mask in real time
✅ Displays class label + confidence
✅ FPS counter (yellow for ~11 FPS on CPU)
✅ Works with multiple faces in frame

---

🛠️ Requirements

• Python 3.8+
• PyTorch
• sklearn (train_test_split only)
• Torchvision
• OpenCV (opencv-python)
• NumPy

---

👤 Author

Built with ❤️ by Youssef Ahmed El Demerdash