Face Recognition using CNN and FaceNet

A deep learning-based CBIR system for face recognition using CNNs, VGG16, and FaceNet, optimized for celebrity face retrieval and classification.

Full, Detailed report and walkthrough is available in Report.pdf

📁 Dataset

Source: Celebrity Face Dataset (Kaggle)
Structure:

• 31 celebrity folders (e.g., Henry Cavill, Natalie Portman)
• 50–150 images per class (varied poses/lighting)
• Cropped faces using MTCNN for focus

🛠️ Methodology

1. Custom CNN

• Architecture:
  Conv2D → MaxPooling → BatchNorm → Dropout → Dense
• Training:
  • 3 classes (Henry Cavill, Natalie Portman, Robert Downey Jr.)
  • 30 epochs, Adam optimizer
  • Accuracy: 81.48% (test set)

2. VGG16 Transfer Learning

• Preprocessing: Data augmentation (rotation, flip, brightness)
• Fine-tuning:
  • Frozen base layers + custom dense layers
  • Accuracy: 92.59% (test set)
• Limitation: Overconfidence in misclassifications

3. FaceNet + SVM

• Embeddings: 128D face features extracted via FaceNet
• Classifier: SVM with linear kernel
• Accuracy: 99.63% (test set)
• Key Advantage: Robust to pose/lighting variations

📊 Key Results

Model	Test Accuracy	F1-Score	Specialization
Custom CNN	81.48%	0.81	Basic face recognition
VGG16	92.59%	0.93	Transfer learning
FaceNet+SVM	99.63%	0.99	State-of-the-art performance

💻 How to Run

1. Install dependencies:

pip install tensorflow keras-facenet scikit-learn mtcnn opencv-python matplotlib