Real-Time Face Recognition with SCRFD + ArcFace (ONNX)

This project implements a modern, high‑accuracy, production‑grade, and fully free face‑recognition pipeline using:

• SCRFD (face detector)
• ArcFace-ResNet100 (face embedding model)
• SVM classifier for personalized recognition
• Real‑time webcam inference

The system is optimized for small datasets, making it ideal for events, access‑control, or personalized recognition tasks.

---

🚀 Features

🔍 Face Detection — SCRFD

• Extremely fast
• Highly accurate
• Robust across lighting, pose, angle, small faces, etc.
• Runs entirely on CPU via ONNX Runtime

🧠 Face Embeddings — ArcFace ResNet100

• State‑of‑the‑art model used in industry
• Produces 512‑dimensional embeddings
• Highly discriminative and stable
• Excellent performance even with 1–3 images per person

🎯 Classifier — SVM

• Works great with small datasets
• No retraining of deep models required
• Very fast training (<1 second typically)

🎥 Real‑Time Recognition

• Uses OpenCV webcam feed
• No caching: each frame is processed fresh
• Works fluently on CPU

---

📦 Project Structure

project/
│
├── models/
│   ├── scrfd_500m_bnkps.onnx
│   ├── scrfd_500m_kps.onnx
│   └── arcfaceresnet100-8.onnx
│
├── person_images/
│   └── person_name/ (one folder per person)
│
├── outputs/
│   └── (embeddings + classifier are saved here)
│
└── Main.py

---

📥 Installing Dependencies

Activate your virtual environment and run:

pip install onnxruntime opencv-python numpy scikit-learn joblib requests

---

📘 How to Use

1️⃣ Place Your Training Images

Inside:

person_images/

Use the structure:

person_images/
   Sourav/
       1.jpg
       2.jpg
       3.jpg
   Alice/
       1.jpg
       2.jpg

You need 1 or more images per person.

---

2️⃣ Run the Script

python Main.py --mode train
python Main.py --mode webcam

What happens:

• SCRFD detects faces in each training image
• ArcFace converts each face → 512‑D embedding
• SVM classifier trains
• Webcam opens
• Each frame is analyzed and recognized live

---

📊 Output Files

Saved in outputs/:

File	Purpose
svm_classifier.pkl	Trained classifier
label_encoder.pkl	Mapping between numeric labels ↔ names
embeddings.npy	Stored embeddings
labels.npy	Stored corresponding labels

---

🎯 Notes for Best Accuracy

• 3–10 photos per person recommended
• Ensure face is clear and well-lit
• Include multiple angles if possible
• Avoid sunglasses or masks

---

🔐 License & Cost

• SCRFD — MIT License
• ArcFace-ResNet100 — Apache 2.0
• ONNX Runtime — MIT
• Everything used is 100% free and requires no subscription.