Siamese Neural Network for Facial Recognition

A Deep Learning application for one-shot face verification using Triplet Loss.

---

Project Overview

This project implements a Siamese Neural Network (SNN) designed to perform facial recognition through embedding similarity, rather than traditional classification. By using Triplet Loss, the network learns to minimize the distance between images of the same person (anchor and positive) while maximizing the distance from a different person (negative).

The model was trained on the Celebrity Faces Dataset, achieving near-perfect verification accuracy on 5-, 10-, 15-, and 20-way tests, and 98% accuracy on 25-way tests — demonstrating robust and generalizable embeddings.

---

Core Objectives

• Build a CNN-based embedding model capable of learning discriminative facial features.
• Train the model using Triplet Loss to improve similarity learning.
• Evaluate model performance across multiple n-way identification tests.
• Visualize embeddings and verify the ability to generalize to unseen faces.

---

Tech Stack

Languages & Frameworks

• Python 3.10
• TensorFlow / Keras
• NumPy | Matplotlib | OpenCV

Development Tools

• Google Colab + GPU Runtime
• VS Code (local testing)
• Git & GitHub (version control)

---

Model Architecture

The embedding model is a sequential CNN comprising:

1. Convolution → BatchNorm → MaxPooling → Dropout (3× blocks)
2. Flatten → Dense(128, ReLU)
3. Lambda Layer for L2 normalization of embeddings

The embeddings for (anchor, positive, negative) images are concatenated to form the Siamese structure.

---

Training Summary

Parameter	Value
Optimizer	Adam
Loss	Triplet Loss
Input Shape	105×105×1 (grayscale)
Embedding Size	128
Epochs	10
Batch Size	4

Training Results

Epoch 1–10: Loss steadily reduced to 8.17e-05

Verification Accuracy

Test Type	Accuracy
5-way	100%
10-way	100%
15-way	100%
20-way	100%
25-way	98%

---

Key Insights

• Triplet Loss directly optimizes similarity relationships, not classification.
• L2-normalized embeddings stabilize cosine distance comparisons.
• Preprocessing (grayscale, resize, normalize) removed noise and ensured consistency.
• CNN feature extractor learned compact, robust facial representations.

---

How to Run

1. Clone the repo:

   git clone https://github.com/JDede1/advanced-machine-learning-applications1.git
   cd advanced-machine-learning-applications1

2. Create and activate your virtual environment:

   python3 -m venv ml-env
   source ml-env/bin/activate   # (Linux/Mac)
   ml-env\Scripts\activate      # (Windows)

3. Install dependencies:

   pip install -r requirements.txt

4. Run the notebook or Python script:

   jupyter notebook notebooks/Facial_recoginition_Project.ipynb
   # or
   python notebooks/facial_recoginition_project.py

---

Business Applications

• Employee or student identity verification systems
• Access control and attendance tracking
• Customer identification in retail / banking
• Face-based authentication for smart devices

---

Recommendations for Production Deployment

• Integrate face detection preprocessing (e.g., MTCNN).
• Use FaceNet-style transfer learning for larger datasets.
• Deploy via FastAPI + TensorFlow Serving for real-time inference.
• Add monitoring (e.g., model drift detection using embeddings).
• Store embeddings securely for privacy compliance.

---

Author

Ajibola Jeremiah Dedenuola University of Arizona — MIS/ML Program Role: Machine Learning Engineer / Data Scientist GitHub Profile