Face Recognition System using Siamese Network (PyTorch + OpenCV)

This is a Python-based Face ID system that can register multiple users, and later verify a person using facial similarity. It uses a Siamese Neural Network implemented in PyTorch and OpenCV for real-time webcam face capture.

---

Features

• Register new users using webcam face capture
• Train a Siamese neural network to learn face similarity
• Verify users live via webcam
• Supports multiple users

What is a Siamese Network?

A Siamese Network is a neural architecture that learns to determine similarity between two inputs rather than classifying them directly. It consists of twin networks that share weights and extract embeddings from input images.

In this project, the network compares two face images:

• If the faces belong to the same person, the distance between embeddings is small.
• If they belong to different people, the distance is large.

It is trained using contrastive loss, which teaches the network to minimize the distance between similar faces and push dissimilar ones apart.

Installation

Install Python dependencies:

pip install opencv-python torch torchvision numpy pillow

Usage Guide

1. Register a User

Run the registration script and follow the prompt:

python register.py

Enter a username (e.g. alice). Press and hold the s key to capture and save face images 300 images will be saved under data/user/alice

2. Train the Model

Train the Siamese network on registered face pairs:

python train.py

Trains a model to learn facial similarity. Model is saved to models/siamese_model.pt

3. Verify a Face

Launch the verification script to recognize a face:

python verify.py

Webcam opens, Press q to capture a face. The system compares it with stored users. Outputs the closest match or shows "Unknown"

Technical Details

Face images are converted to grayscale and resized to 100x100 The model uses contrastive loss to learn embedding space During verification, it uses pairwise distance between embeddings Adjustable verification threshold (default 0.8) in verify.py: THRESHOLD = 0.8

Requirements

Python 3.7 or later PyTorch torchvision OpenCV (opencv-python) Pillow NumPy

$ python register.py Enter username to register: bob Capturing 10 face images for user: bob Saved image 1/10 ...

$ python train.py Epoch 1/10, Loss: 0.3912 ... Model saved to models/siamese_model.pt

$ python verify.py Best match: bob (score: 0.5473) Verified as: bob

To-Do (Optional Ideas)

Use MTCNN or Dlib for better face detection Cache face embeddings for faster verification Add feature for deleting/updating users GUI version (Tkinter or PyQt)