Classical Face Biometrics

An educational collection of reproducible experiments in biometric verification and classical face recognition. The repository implements the core numerical steps explicitly with NumPy, making each method compact enough to study while still producing publication-ready plots, CSV summaries, and reusable projection artifacts.

The project covers score-based verification, Eigenfaces, Fisherfaces, illumination normalization, and local patch descriptors.

Highlights

• Explicit implementations of ROC metrics, PCA, LDA, KNN, integral images, and patch-level voting.
• Deterministic dataset protocols and validation of the expected directory structure.
• A single command that regenerates every table, model artifact, and figure.
• Focused unit tests for the numerical building blocks.
• Generated results are versioned, so the repository remains browsable without access to the original image datasets.

Experiments

Experiment	Method	Dataset / input	Main output
Performance metrics	ROC, AUC, EER, d-prime	Client and impostor scores	Verification operating points
Eigenfaces	PCA + KNN	ORL	Error across PCA dimensions
Fisherfaces	PCA + LDA + KNN	ORL	Fisherfaces vs Eigenfaces
Local normalization	Local statistics + CLAHE	Extended Yale B	Illumination-distance analysis
Local face features	Patch PCA + local voting	ORL	Patch and window-size sweeps

Verified Results

All experiments were regenerated on June 12, 2026 with Python 3.11.7.

Experiment	Best or representative result
Verification system A	AUC 0.8832, EER 0.2013, d-prime 1.0747
Verification system B	AUC 0.8831, EER 0.2010, d-prime 1.2353
Eigenfaces	89.5% accuracy with d'=80, k=1
Fisherfaces	85.5% accuracy with d'=39, k=1
Local normalization	Mean cosine distance reduced from 0.1210 to 0.0696 with window 35
Local features	56.9% accuracy with sequential patches, window 16, and 12 patches

These values describe the fixed protocols implemented in this repository; they should not be interpreted as general benchmarks for the underlying datasets.

Example Outputs

Eigenfaces	Fisherfaces

Illumination Normalization

Verification Performance

Installation

Python 3.11 is recommended.

git clone https://github.com/JordiCan/biometry.git
cd biometry
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

The main dependencies are NumPy, SciPy, Matplotlib, Pillow, and scikit-image. Matplotlib uses a non-interactive backend, so the experiments also run on headless systems and in CI environments.

Dataset Setup

The image datasets are intentionally excluded from Git because of their size and licensing terms. Obtain them from their official or authorized source and place them under data/ with this layout:

data/
├── IDENTIFICACION/
│   └── ORLProcessed/
│       ├── Train/
│       │   ├── s1/1.pgm ... 5.pgm
│       │   └── ...
│       └── Test/
│           ├── s1/6.pgm ... 10.pgm
│           └── ...
└── ILLUMINATION/
    └── YaleB/
        ├── yaleB02/*.pgm
        ├── yaleB03/*.pgm
        └── ...

The verification score files are small and are included in performance_metrics/data/. The ORL scripts expect 40 subjects and the fixed 1-5 train / 6-10 test protocol. The local-feature experiment uses images 1-2 for training and 3-10 for testing. The illumination experiment selects 12 Yale B subjects and 24 images per subject.

Reproduce Everything

Run every experiment and auxiliary plot generator in a stable order:

python run_all.py

On the reference environment the full run takes roughly two minutes. Each script can also be executed independently:

python performance_metrics/evaluate.py
python performance_metrics/evaluate_heaviside.py
python eigenfaces/eigenfaces.py
python eigenfaces/aux.py
python fisherfaces/fisherfaces.py
python local_face_normalization/local_face_normalization.py
python local_face_normalization/aux.py
python local_face_features/local_face_features.py

Outputs are written to each experiment's plots/ directory. Eigenfaces and Fisherfaces additionally save learned means, projection matrices, labels, and projected vectors under outputs/.

Testing

The unit tests do not require the image datasets:

python -m unittest discover -s tests -v

GitHub Actions runs the same tests on Python 3.10, 3.11, and 3.12.

Repository Layout

.
├── performance_metrics/       # Verification metrics and score plots
├── eigenfaces/                # PCA face recognition
├── fisherfaces/               # PCA + LDA face recognition
├── local_face_normalization/  # Illumination preprocessing
├── local_face_features/       # Patch descriptors and local voting
├── tests/                     # Dataset-independent regression tests
├── run_all.py                 # Full reproduction entry point
├── requirements.txt
└── LICENSE

Each experiment directory has its own README with method details, generated figures, output descriptions, and an interpretation of the results.

Reproducibility Notes

• The experiments contain no random sampling; repeated runs with the same dependency versions and datasets are deterministic.
• CSV values are rounded for presentation, while calculations use float64.
• AUC uses pairwise client/impostor comparisons with half credit for ties.
• The threshold convention is score >= threshold.
• d-prime uses the pooled population variance: (mean_client - mean_impostor) / sqrt((var_client + var_impostor) / 2).
• Generated binary arrays can vary at the last floating-point digits across BLAS/LAPACK implementations while preserving the reported metrics.

License

The source code is available under the MIT License. Dataset licenses remain with their respective owners and are not covered by this repository's license.