Partial Face Recognition with Deep Learning

This final year undergraduate project explores facial recognition under real-world constraints by developing a deep learning system capable of identifying partially visible faces. Built on a fine-tuned VGG16 architecture, the system was trained on the LFW and PartialLFW datasets using transfer learning, facial alignment, and data augmentation to improve robustness against occlusions such as masks, sunglasses, and off-angle poses.

The project was completed as part of a BSc in Computing (Cybersecurity) at the University of Buckingham under the supervision of Dr. Naseer Al-Jawad.

---

📌 Key Highlights

• Developed a custom face recognition pipeline capable of handling partial occlusions
• Fine-tuned a pre-trained VGG16 model with TensorFlow/Keras
• Applied transfer learning and replaced fully connected layers for task-specific adaptation
• Integrated Dlib facial landmark detection and OpenCV preprocessing
• Trained on both LFW and PartialLFW datasets
• Achieved 50.16% accuracy on occluded faces — with clear potential for further refinement
• Tackled technical challenges including hardware limitations, shape mismatches, and environment setup

---

🧠 Technical Stack

• Python 3.8
• TensorFlow + Keras
• OpenCV + Dlib for alignment and preprocessing
• NumPy / Pandas / Scikit-learn for data manipulation and evaluation

---

📊 Dataset Summary

• Labelled Faces in the Wild (LFW) – standard facial recognition dataset
• PartialLFW – enhanced dataset with occlusions (e.g. masks, glasses, angles)

---

🧪 Methodology Overview

1. Face Alignment using Dlib landmarks
2. Preprocessing to resize (224x224), normalize, and augment images
3. Transfer Learning from ImageNet-pretrained VGG16
4. Model Fine-Tuning and classification via custom fully connected layers
5. Training with Early Stopping and checkpointing
6. Evaluation using accuracy, F1 score, and validation graphs

---

📁 Repository Structure

Project-partial-face-recognition/ ├── align/ # Alignment scripts and resources ├── input_folder/ # Sample images ├── partial_faces/ # Region-extracted data (eyes, nose, etc.) ├── scripts/ │ ├── align_dataset.py │ ├── gen_partial_faces.py │ ├── preprocess_images.ipynb │ ├── split_dataset.ipynb │ └── training_vgg16.ipynb ├── README.md ├── requirements.txt └── final_report.pdf # Full academic write-up

---

🚀 Getting Started

pip install -r requirements.txt
python scripts/align_dataset.py
python scripts/gen_partial_faces.py
jupyter notebook scripts/training_vgg16.ipynb

🎓 Academic Context
Project Title: Partial Face Recognition 
Author: Julia Szostakiewicz
Degree: BSc (Hons) Computing (Cybersecurity)
Award: First Class Honours
Institution: University of Buckingham
Supervisor: Dr Naseer Al-Jawad