FaceReg - Real-Time Face Recognition System

A face recognition application built with MTCNN + InceptionResnetV1 (FaceNet) and deployed via Streamlit. Supports image, video, and real-time face recognition with dynamic face enrollment.

Architecture

Input (Image / Video / Camera)
        |
        v
  +-----------+       +-------------------+       +------------------+
  |   MTCNN   | ----> | InceptionResnetV1 | ----> | Euclidean Match  |
  | Detection |       |  512-d Embedding  |       | Threshold = 2.0  |
  +-----------+       +-------------------+       +------------------+
        |                                                |
   Bounding Box                                    Name + Confidence

Features

Feature	Description
Image Recognition	Upload an image and identify faces
Video Recognition	Process MP4 videos with face detection overlay
Real-Time Recognition	Live camera feed recognition
Face Enrollment	Train new faces via video upload or webcam
Unknown Detection	Identifies unknown persons (distance > 2.0)

Tech Stack

• Deep Learning: PyTorch, FaceNet-PyTorch (MTCNN + InceptionResnetV1)
• Web Framework: Streamlit, Streamlit-WebRTC
• Computer Vision: OpenCV
• ML Evaluation: Scikit-learn (SVM comparison in notebooks)
• Deployment: Docker

Demo & Results

Processing Pipeline

Image → MTCNN Detection → InceptionResnetV1 Encoding → Euclidean Matching → Output

Recognition Results

Image Recognition	Video Recognition
Upload image → detect & identify face	Process video frames with bounding box overlay

Performance

Metric	Value
Accuracy	97% on 31-class celebrity dataset
Embedding Dim	512-d vectors
Pre-training	VGGFace2
Train/Test Split	70% / 30%
Dataset	2,562 images across 31 classes (Kaggle)

Project Structure

FaceReg/
├── app.py                          # Entry point
├── src/
│   ├── core/
│   │   └── face_engine.py          # Face detection, encoding & matching
│   └── ui/
│       ├── styles.py               # Custom CSS styling
│       └── pages.py                # Streamlit page functions
├── models/
│   ├── vector_face.pth             # Pre-trained face embeddings
│   └── name_face.npy               # Face labels (31 classes)
├── notebooks/
│   ├── euclidean_distance.ipynb    # Euclidean distance approach
│   └── svm_classifier.ipynb        # SVM classifier approach
├── assets/                          # Static files (images, gifs)
├── experiments/
│   └── app_webrtc.py               # WebRTC experimental version
├── Dockerfile
├── requirements.txt
└── README.md

Quick Start

# Clone the repository
git clone https://github.com/HieuNTg/FaceReg.git
cd FaceReg

# Install dependencies
pip install -r requirements.txt

# Run the app
streamlit run app.py

Docker

docker build -t facereg .
docker run -p 8501:8501 facereg

Open http://localhost:8501 in your browser.

Training Notebooks

Notebook	Approach	Details
notebooks/euclidean_distance.ipynb	Euclidean Distance	Distance-based matching with averaged embeddings
notebooks/svm_classifier.ipynb	SVM Classifier	Support Vector Machine on 512-d embeddings

Future Improvements

• Liveness Detection - Anti-spoofing with face mesh / blink detection
• Multi-Face Recognition - Detect and identify all faces in a frame simultaneously
• Cosine Similarity - Alternative distance metric for better embedding comparison
• Attendance System - Automated check-in/check-out with reporting
• Edge Deployment - ONNX/TensorRT conversion for Raspberry Pi / Jetson Nano
• Vector Database - Milvus/Pinecone for scalable face search