RetinaFace Masked Face Detection & Mask Classifier

RetinaFace + MobileNet Mask Classifier — Training, Testing & Live Recognition Pipeline

---

🌐 Overview

This repository contains a training and inference pipeline built around RetinaFace (for face detection) and a MobileNetV2-based Mask Classifier trained from scratch on curated datasets (RMFD, MAFA, CMFD, and custom images).
It powers a real-time masked face detection system.

---

🧠 Features

• RetinaFace for robust face detection (supports ResNet50 & MobileNet backbones)
• MobileNetV2 classifier trained for Mask vs No Mask
• 99.3% test accuracy on curated test set
• Supports real-time webcam inference
• Seamlessly integrates with facial recognition backend
• Modular & extensible for future fine-tuning

---

⚙️ Installation

# Create & activate virtual environment
python -m venv venv_train
venv_train\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

Use the CUDA-enabled PyTorch build matching your GPU.

---

🧾 Dataset Setup

The Dataset is prepared and filtered from various real World Datasets like RMFD,CMFD,MAFA, ANd Custom Real Masked images. Structure your dataset as follows:

data/Dataset/
├── train/
│   ├── Mask/
│   └── No_mask/
├── val/
│   ├── Mask/
│   └── No_mask/
└── test/
    ├── Mask/
    └── No_mask/

Ensure a balanced dataset (~9k Mask / 9k No Mask).
Supports custom datasets for real-world adaptation.

---

🏋️‍♂️ Training the Classifier

python train.py

train.py uses:

• MobileNetV2 base network
• Adam optimizer
• CrossEntropy loss
• Early stopping based on validation accuracy

Checkpoints are saved in checkpoints/ as:

• mobilenet_mask_best.pth.tar
• mobilenet_mask_last.pth.tar

Training curves

---

🔬 Testing

All The Trained weights are available here.

users must have Git LFS installed to download large weights. Evaluate your trained model:

python test_script.py

Example output:

====== TEST REPORT ======
Total images: 1849
Mask: 924 (49.97%)
No Mask: 925 (50.03%)
Skipped: 0 (0.00%)
==========================
Overall Accuracy: 99.3%

---

🎥 Live Detection Demo

Run real-time webcam detection:

python mask_detector_webcam_test.py

• Detects multiple faces in real time
• Classifies mask status using your trained MobileNet
• Overlays bounding boxes with confidence scores

Example result:

| | |

---

🧩 Model Integration

The trained model integrates seamlessly with the Face Recognition System, where:

1. RetinaFace detects faces in real-time video or CCTV frames.
2. The MobileNet classifier determines if the face is masked.
3. Embeddings are generated.
4. Matches are found from the Database.

---

📊 Performance Summary

Component	Backbone	Accuracy	FPS	Notes
RetinaFace Detector	MobileNet0.25	97.8%	30	Fast & lightweight
Mask Classifier(ours)	MobileNetV2	99.3%	28	Custom trained
Combined (End-to-End)	RetinaFace + MobileNet	98.9%	25	Optimized for real-time

---

---

📎 Why Custom Training?

While public repos like [Face-Mask-Detection] exist,
this model was custom-trained because:

• It uses real forensic and surveillance-grade images.
• Integrated mask detection within the recognition flow.
• Optimized for speed and accuracy balance with MobileNet.
• Full control over architecture, logging, and checkpoints.

---

🧾 License & Acknowledgements

• Licensed under the MIT License.
• Based on the open-source RetinaFace project.
• Datasets used: RMFD, MAFA, CMFD, Custom Surveillance Dataset.
• All the images and datasets and models used here belongs to the respective owners.
• MIT © Joel Thomas

---

Code of Conduct

You can find our Code of Conduct here.

👤 Author & Contact

👨‍💻 Joel Thomas

• 🔗 LinkedIn
• 💻 GitHub
• 📧 Email: [email protected]

---

⭐ Contribute

Pull requests and issues are welcome.
You can contribute by improving dataset balance scripts, fine-tuning on other backbones, or optimizing for embedded systems.

---

“Built with purpose — precision and performance for real-world recognition.”