Deep Learning Facial Recognition (Python)

A compact, educational demo that implements facial verification using a Siamese neural network. The main workflow is provided as a Jupyter notebook (deep_learning_facial_recognition.ipynb) that walks through image collection, preprocessing, model training, and a simple verification routine using OpenCV.

This repository is intended for learning and experimentation only. It demonstrates concepts (pairwise learning, L1 distance layer, webcam capture) and is not intended as a production-ready facial recognition system.

Quick links

• Notebook: deep_learning_facial_recognition.ipynb
• Example inference script: predict.py
• Requirements: requirements.txt

Features

• Webcam-based image collection helpers (anchor / positive / negative)
• Data preprocessing and dataset construction using TensorFlow
• Compact Siamese architecture with a custom L1 distance layer
• Training loop with checkpointing and a lightweight inference helper

Quickstart (Windows PowerShell)

1. Clone the repository and change into it:

git clone https://github.com/royxlead/deep-learning-facial-recognition-python.git
cd deep-learning-facial-recognition-python

2. Create and activate a virtual environment, then install dependencies:

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

3. (Optional) Create example directories the notebook expects:

python -c "import os; [os.makedirs(p, exist_ok=True) for p in ['data/anchor','data/positive','data/negative','application_data/input_image','application_data/verification_images']]"

4. Start Jupyter and open the notebook:

jupyter notebook deep_learning_facial_recognition.ipynb

Run the top-of-notebook setup cell first (it checks the environment, GPU availability, and sets seeds), then run cells in order.

Running inference (example)

After training and saving a model (the notebook saves siamese_model.keras by default), you can run the provided predict.py script:

python predict.py --input application_data/input_image/input_image.jpg --verification_dir application_data/verification_images --model siamese_model.keras --threshold 0.5

The script prints per-image similarity scores and a final verification ratio.

Project layout

deep-learning-facial-recognition-python/
├── deep_learning_facial_recognition.ipynb   # Notebook (main workflow)
├── predict.py                               # Minimal inference wrapper
├── requirements.txt                          # Pinned dependency list
├── .gitignore
├── data/                                     # training data (anchor/positive/negative) - NOT committed
└── application_data/                         # input/verification images - NOT committed

Notes & recommendations

• The notebook contains simplified preprocessing and a small model for educational use. For better results consider:

  • performing face alignment and cropping using a robust face detector
  • increasing dataset size and variety
  • replacing the toy embedding network with a pretrained backbone (MobileNet/ResNet) and fine-tuning

• Security & privacy: collecting and using face images has legal and ethical implications. Only use images you have permission to use and follow local laws and best practices.

Contributing

Contributions and improvements are welcome. Suggested small first PRs:

• add a GitHub Actions workflow that runs a quick smoke test on predict.py (use CPU-only TensorFlow and small sample images)
• parameterize notebook paths and hyperparameters for non-interactive runs
• add an example dataset or scripts to download a small public sample

When opening PRs, include a short description and a smoke-test showing the change works.

License

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

Acknowledgements

• TensorFlow — model building and training
• OpenCV — webcam capture and visualization

---

If you want, I can add a small sample image and a CI workflow that runs predict.py as a smoke test (CPU-only). Which would you prefer next?