A facial verification system using a Siamese Neural Network architecture built with PyTorch. This project enables real-time facial recognition with a webcam to verify if a person matches registered faces in the verification database.

Verified (Hi 3llam 🫡) ✅	Unverified (Not Recognized ❌)
System recognizes the registered face	System doesn't recognize the face

This system uses one-shot learning via a Siamese network to perform facial verification without requiring large datasets of the target individual. The model compares an input face against a database of verification images and determines if it's a match based on configurable thresholds.

• Siamese Neural Network 🔄: Utilizes twin networks with shared weights for facial comparison
• One-Shot Learning 📸: Requires only a small set of reference images
• Real-time Verification ⚡: Process webcam feed for instant facial verification
• Customizable Thresholds 🎛️: Adjustable detection and verification thresholds
• Desktop Application 🖥️: Built with Kivy for cross-platform GUI interface

Deep-Facial-Recognition/
├── app/                        # Main application folder
│   ├── __init__.py
│   ├── faceid.py               # Kivy application for real-time verification
│   ├── model.py                # Neural network model architecture
│   ├── siamese_network_v2.pth  # Trained model weights
│   └── application_data/       # Application data directory
│       ├── input_image/        # Stores input image from webcam
│       └── verification_images/ # Reference images to compare against
├── application_data/           # Root application data (mirror of app/application_data)
├── data/                       # Training data directory
│   ├── anchor/                 # Anchor images (training)
│   ├── positive/               # Positive images (training)
│   └── negative/               # Negative images (training)
├── models/                     # Stored model weights
├── training_checkpoints/       # Training checkpoint files
└── Facial Verification with a Siamese Network.ipynb  # Training notebook

The system implements a Siamese Neural Network with:

• Embedding network with CNN layers for feature extraction 🧩
• L1 Distance layer to compute similarity between embeddings 📏
• Classification layer to determine verification probability 📊

The model was trained in two distinct phases:

Phase 1 📊:

• Training on 300 image samples with no augmentation
• Resulted in initial model (siamese_network.pth)
• Established baseline performance

Phase 2 📈:

• Training on 3000 image samples with data augmentation
• Applied transformations including random flips, brightness/contrast adjustments
• Produced the final optimized model (siamese_network_v2.pth)
• Significant improvement in accuracy and generalization

The model was trained using triplets of:

• Anchor images (reference face) ⚓
• Positive images (same person as anchor) ✅
• Negative images (different person) ❌

The model learns to minimize the distance between anchor-positive pairs while maximizing the distance between anchor-negative pairs.

• Python 3.8+ 🐍
• PyTorch 🔥
• OpenCV 👁️
• Kivy (for the desktop GUI application) 🖼️
• PIL (Pillow) 🖼️

1. Clone the repository:

git clone https://github.com/Ad7amstein/Deep-Facial-Recognition.git
cd Deep-Facial-Recognition

2. Install dependencies:

pip install -r requirements.txt

3. Running the application:

cd app
python faceid.py

1. Launch the desktop application 🖥️
2. Position your face in the camera frame 😊
3. Click "Verify" button to check if your face matches the verification database 🔍
4. Result will be displayed on screen ✓

The complete training process is documented in the Jupyter notebook. To train with your own data:

1. Collect images for the anchor, positive, and negative directories 📸
2. Run the notebook cells to:
   • Preprocess images 🔄
   • Create the dataset 📦
   • Train the Siamese network 🏋️‍♂️
   • Evaluate and save the model 💾

This project is licensed under the MIT License - see the LICENSE file for details.

• The Labeled Faces in the Wild dataset was used for negative samples during training
• This project implements concepts from the paper Siamese Neural Networks for One-shot Image Recognition by Gregory Koch, Richard Zemel, and Ruslan Salakhutdinov (2015)