Face Authentication Using Classical Image Processing

A face authentication system using traditional image processing techniques.

---

Overview

This project was developed as the final project for the "Introduction to Image Processing" course at National Dong Hwa University (NDHU). The goal was to implement a face authentication system using classical image processing techniques without relying on deep learning models.

It uses facial landmark detection, alignment, and feature normalization to compare facial features based on Euclidean distance.

Features

• Face alignment based on eye landmarks
• Symmetric feature adjustment for consistent representation
• Euclidean distance-based face matching
• Threshold optimization to improve recognition performance
• Performance evaluation using accuracy, precision, recall, and F1-score

---

Dataset

The dataset consists of two main directories:

Face_DB/ (Authorized faces)

• Images/ → Frontal face images (e.g., NAME_000.jpg)
• Landmark_data/ → Corresponding landmark CSV files

Test_DB/ (Test images)

• Images/ → Face images at different angles
• Landmark_data/ → Landmark CSVs for each test image

Each person has 6 test images, totaling 90 test samples.

---

Methodology

1️. Face Alignment Using Eye Landmarks

• Faces were aligned using eye landmarks to ensure consistency.

2️. Symmetrization of Facial Features

• Each landmark was mirrored and adjusted to reduce asymmetry caused by angle variations.

3️. Feature Extraction & Normalization

• Jawline excluded to prevent distortions.
• Relative distances between landmarks used instead of absolute positions.
• Feature distances were normalized using the width of the left eye.

4️. Face Matching Using Euclidean Distance

• Euclidean distance was computed between the aligned and normalized feature vectors.

5️. Threshold Selection for Best Performance

• Thresholds were tested between 50 and 120 to optimize results.
• Best threshold: 73 (Selected based on F1-score).

---

Results

Metric	Value
Accuracy	52.22%
Precision	57.69%
Recall	58.82%
F1-score	58.25%

Example Predictions

Scenario	Test Image	Predicted Person	Verdict
✅ Correct Match	Chloe_001.jpg	Chloe	✔️ Accepted
❌ False Positive	Ethan_002.jpg	Sebastian	❌ Imposter Accepted
❌ Wrong Identity	Zoe_000.jpg	Chloe	❌ Misclassified
✅ Correct Rejection	Lucas_004.jpg	—	✔️ Rejected
❌ False Rejection	Wyatt_004.jpg	—	❌ Incorrectly Rejected

---

How to Run the Project

1️. Clone the Repository

git clone https://github.com/chitsip149/face-recognition.git
cd "411021365 Nguyen Minh Trang"

2️. Install Dependencies

pip install -r requirements.txt

3️. Run the Face Authentication System

python face_auth.py

4️. (Optional) Run Performance Evaluation

python evaluate.py

Key Learnings

• Preprocessing (alignment & normalization) is critical in face recogition.
• Feature selection (excluding jawline) improves robustness to angles.
• Euclidean distance works for simple comparisons but has limitations.
• Empirical evaluation helps fine-tune parameters like the threshold.

Future Improvements

• Use machine learning models (e.g., SVM, k-NN) for improved classification.
• Apply deep learning (CNNs, OpenCV DNN, or FaceNet) for robust face recognition.
• Optimizr featue selection using principal component analysis (PCA).
• Expand dataset to test on more diverse facial variations.

License

This project is licensed under the MIT License.