🎭 Face Recognition System

A dual-model face recognition system built with Python, supporting both a custom CNN (PyTorch) and a classical LBPH (OpenCV) approach for real-time face identification via webcam.

---

📋 Table of Contents

• Overview
• Project Structure
• Models
• Installation
• Usage
  • CNN Pipeline
  • LBPH Pipeline
• Configuration
• How It Works

---

Overview

This project provides two independent face recognition pipelines:

Feature	CNN Model	LBPH Model
Backend	PyTorch	OpenCV
Speed	GPU-accelerated (if available)	CPU-only
Accuracy	Higher (deep learning)	Moderate (classical ML)
Training time	Longer	Very fast
Confidence threshold	90% softmax probability	≤ 60 LBPH distance

---

Project Structure

face_recognition/
├── config.py               # Centralized configuration (paths, hyperparameters)
├── model.py                # CNN model architecture definition
├── dataCollection.py       # Webcam-based face image collection (for CNN)
├── trainer.py              # CNN model training + confusion matrix evaluation
├── main.py                 # Real-time CNN face recognition (webcam)
├── LBPH_model.py           # All-in-one LBPH pipeline (collect → train → recognize)
├── face_recognition_cnn.pth  # Saved CNN model weights (generated after training)
├── trainer.yml             # Saved LBPH recognizer (generated after training)
└── data/
    └── train/
        ├── Person_A/       # One folder per person
        │   ├── 1.jpg
        │   └── ...
        └── Person_B/
            └── ...

---

Models

🧠 CNN Model (model.py)

A custom 3-layer Convolutional Neural Network:

• Conv layers: 3×3 kernels, ReLU activation, channels: 3 → 32 → 64 → 128
• Pooling: MaxPool2d (×2) after conv2 and conv3, halving spatial dims twice
• FC layers: Flattened → 128 → num_classes (raw logits for CrossEntropyLoss)
• Input size: Configurable via IMAGE_SIZE (default 300×300)

📊 LBPH Recognizer (LBPH_model.py)

Uses OpenCV's built-in LBPHFaceRecognizer on grayscale, 200×200 face crops detected with a Haar Cascade.

---

Installation

Prerequisites

• Python 3.8+
• Webcam

Install Dependencies

pip install opencv-python opencv-contrib-python torch torchvision pillow matplotlib seaborn scikit-learn numpy

GPU Support: Install the CUDA-compatible version of PyTorch from pytorch.org for faster CNN training/inference.

---

Usage

CNN Pipeline (main.py / trainer.py / dataCollection.py)

Step 1 — Collect Training Data

python dataCollection.py

• Enter the person's name when prompted.
• A webcam feed opens with live face detection (Haar Cascade).
• Press s to save a cropped face image.
• Press q or 0 to quit. Up to 2000 images per person.
• Images are saved to data/train/<name>/.

Step 2 — Train the CNN

python trainer.py

• Loads all images from data/train/, splits into train/unused (configurable via TRAIN_SPLIT).
• Trains for NUM_EPOCHS epochs using Adam optimizer + CrossEntropyLoss.
• Saves model weights to face_recognition_cnn.pth.
• Displays a confusion matrix heatmap after training.

Step 3 — Run Real-Time Recognition

python main.py

• Loads the saved .pth model.
• Opens webcam; detects faces via Haar Cascade.
• Classifies each face with the CNN — labels it with a name if confidence > 90%, otherwise "Unknown".
• Press q to quit.

---

LBPH Pipeline (LBPH_model.py)

A self-contained script with an interactive menu:

python LBPH_model.py

1: Collect Dataset   → Capture up to 500 grayscale face images per person
2: Train Recognizer  → Train LBPH and save to trainer.yml
3: Live Recognition  → Recognize faces in real-time from webcam
4: Exit

• Collection: Press c to capture, ESC to stop.
• Recognition: Labels a face as "Unknown" if LBPH confidence > 60 (higher = less certain in LBPH).

---

Configuration

All key parameters live in config.py:

Variable	Default	Description
DATA_PATH	data/train	Root dataset directory
MODEL_PATH	face_recognition_cnn.pth	CNN model save path
TRAINER_YML_PATH	trainer.yml	LBPH model save path
HAARCASCADE_PATH	haarcascade_frontalface_default.xml	Haar Cascade file
IMAGE_SIZE	(300, 300)	Input size for CNN
BATCH_SIZE	32	Training batch size
LEARNING_RATE	0.001	Adam optimizer LR
NUM_EPOCHS	2	Training epochs
TRAIN_SPLIT	0.5	Fraction of data used for training
DEVICE	auto (CUDA / CPU)	PyTorch compute device
DETECTION_CONFIDENCE	0.7	Face detection threshold
OFFSET	20	Pixel padding around detected faces

---

How It Works

Webcam Frame
     │
     ▼
Haar Cascade          ← Detects face bounding boxes
     │
     ▼
Face Crop + Resize    ← Normalized to fixed resolution
     │
     ├─── CNN Path ──► PyTorch Model ──► Softmax ──► Name (if conf > 90%)
     │
     └─ LBPH Path ──► LBPH Predict  ──► Name (if dist ≤ 60)

Both pipelines use OpenCV's Haar Cascade (haarcascade_frontalface_default.xml) bundled with OpenCV for face detection, then hand off the cropped region to their respective recognizer.