🧠 Face Re-Identification Using Deep Learning

This project implements a Face Re-Identification System to recognize individuals based on facial features using deep learning. Three state-of-the-art CNN models were trained and compared for accuracy, speed, and deployment readiness.

---

📌 Project Objective

To classify faces of known individuals—even when the images are altered via augmentation—using MobileNetV2, VGG16, and ResNet50V2. This system can be extended to surveillance, biometric verification, or access control.

---

🧾 Dataset Information

• Total Individuals (Classes):

  1. Chakri
  2. Sai Priya
  3. Uday
  4. Vijay
  5. Kavya Sri
  6. Gowthami
  7. Kumar

• Total Images: 31,976

• Images per Class (after augmentation): ~4000

• Train/Val/Test Split: 80% / 10% / 10%

📈 Augmentations Applied

• Gaussian Blur
• Grayscale Conversion
• Horizontal Flip

---

🧠 Models Used

Model	Description
MobileNetV2	Lightweight, mobile-friendly CNN
VGG16	Classic deep CNN for image classification
ResNet50V2	Residual network with skip connections

---

⚙️ Training Configuration

• Batch Size: 32
• Epochs: 10 (Early Stopping enabled)
• Optimizer: Adam
• Loss Function: Categorical Crossentropy

📉 Best Epochs & Validation Loss

Model	Best Epoch	Validation Loss
MobileNetV2	3	1.53
ResNet50V2	3	0.051
VGG16	2	0.19

---

🚀 Inference Speed Comparison

🔧 Keras (.h5 format)

Model	Size	10 Images	50 Images	100 Images
MobileNetV2	146 MB	18.75 s	58.72 s	117.18 s
VGG16	216 MB	5.11 s	19.62 s	39.62 s
ResNet50V2	282 MB	11.01 s	12.85 s	25.78 s

🪄 ONNX (.onnx format)

Model	Size	10 Images	50 Images	100 Images
MobileNetV2	48.4 MB	0.27 s	1.23 s	1.73 s
VGG16	72.1 MB	1.98 s	9.14 s	18.18 s
ResNet50V2	153 MB	0.83 s	3.18 s	6.21 s

---

🧪 Sample Predictions

✅ MobileNetV2

Predicted Class	Actual Class
Uday	Chakri
Chakri	Chakri
Gowthami	Gowthami
Uday	Kavya

✅ VGG16

Predicted Class	Actual Class
Chakri	Chakri
Gowthami	Gowthami
Kavya	Kavya

✅ ResNet50V2

Predicted Class	Actual Class
Chakri	Chakri
Gowthami	Gowthami
Kavya	Kavya

---

✅ Summary

• Best Overall Accuracy: ResNet50V2
• Fastest Inference (ONNX): MobileNetV2
• Best for Mobile Deployment: MobileNetV2
• Most Accurate: ResNet50V2 (low validation loss)

This project demonstrates how model choice, deployment format, and preprocessing affect performance in face re-identification tasks.

---

📦 Requirements

Install dependencies using:

pip install numpy pandas opencv-python tensorflow keras scikit-learn matplotlib onnx onnxruntime