https://github.com/ebrahimgolriz/face-recognition-with-cnn

Developed a convolutional neural network (CNN) utilizing FaceNet to detect and identify multiple faces in an image with high accuracy.
https://github.com/ebrahimgolriz/face-recognition-with-cnn

cnn deep-learning face-recognition facenet keras support-vector-machines vgg16

Last synced: 5 months ago
JSON representation

Developed a convolutional neural network (CNN) utilizing FaceNet to detect and identify multiple faces in an image with high accuracy.

• Host: GitHub
• URL: https://github.com/ebrahimgolriz/face-recognition-with-cnn
• Owner: EbrahimGolriz
• Created: 2025-02-08T22:51:43.000Z (5 months ago)
• Default Branch: main
• Last Pushed: 2025-02-08T23:18:21.000Z (5 months ago)
• Last Synced: 2025-02-09T00:20:10.101Z (5 months ago)
• Topics: cnn, deep-learning, face-recognition, facenet, keras, support-vector-machines, vgg16
• Language: Python
• Homepage:
• Size: 6.84 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Face Recognition using CNN and FaceNet

![Python](https://img.shields.io/badge/Python-3.8%2B-blue)

![Deep Learning](https://img.shields.io/badge/Framework-TensorFlow%2FKeras-orange)

![Accuracy](https://img.shields.io/badge/Accuracy-99.63%25-brightgreen)

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)](https://www.kaggle.com/datasets/vasukipatel/face-recognition-dataset)  

**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:

```bash

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