Facial Verification App

This project is a real-time facial verification system built with Kivy and TensorFlow, utilising a Siamese Neural Network for secure and reliable identity verification via webcam. This app compares a live input image against pre-stored verification images to determine identity matches, making it useful for access control, secure authentication, or biometric experiments.

Table of Contents

• Project Description
• Installation
• Usage
• Tests
• Credits
• License

---

Project Description

This application captures a user's webcam image and compares it to a set of stored reference images using a pre-trained Siamese model. The model evaluates the similarity between two faces using a custom L1 distance layer and makes a verification decision based on defined thresholds.

The app is implemented using:

• Kivy: for real-time GUI and webcam interaction
• TensorFlow/Keras: for deep learning inference
• OpenCV: for image processing and capture
• NumPy: for tensor manipulation and aggregation

---

Installation

Requirements

• Python 3.6–3.10
• pip
• Virtual environment (recommended)

Dependencies

pip install kivy tensorflow opencv-python numpy

Clone the Repository

git clone https://github.com/afras23/facial-verification-app.git
cd facial-verification-app

File Structure

faceverapp/
│
├── app/
├── application_data/
│   ├── input_image/
│   └── verification_images/
├── training_checkpoints/
├── siamesemodelv2.keras
├── faceid.py
├── layers.py
├── Facial_Verification_with_Siamese_Network.ipynb
└── README.md

---

Usage

1. Prepare reference images: Add 1 or more JPEG images to application_data/verification_images/.

2. Run the app:

python faceid.py

3. Verify your face:
   • The app opens a window with a webcam feed.
   • Click the Verify button.
   • Your face is compared against the stored images.
   • Verification result: Verified or Unverified.

---

Tests

Manual Test Workflow

1. Add a valid image to application_data/verification_images/.
2. Run the app, ensure webcam captures and resizes correctly.
3. Test correct predictions with both matching and non-matching faces.

Example Unit Test (in test_layers.py)

import tensorflow as tf
from layers import L1Dist

def test_l1_distance():
    l1 = L1Dist()
    a = tf.constant([[1.0, 2.0, 3.0]])
    b = tf.constant([[2.0, 2.0, 4.0]])
    result = l1(a, b).numpy()
    assert all(result == [1.0, 0.0, 1.0]), f"Unexpected output: {result}"

if __name__ == "__main__":
    test_l1_distance()
    print("L1Dist test passed!")

Run the Tests

python test_layers.py

---

Credits

• Author: Anesah Fraser
• Siamese network model architecture and application inspired by modern facial recognition systems, particularly the paper “Siamese Neural Networks for One-shot Image Recognition” by Koch et al., which guided the development of the model.