Face Recognition Custom CNN with PyTorch

Real-time pipeline for analyzing and recognizing faces using a Custom Convolutional Neural Network (CNN) in PyTorch — with automated aspect-ratio preserving data collection, Haar cascade detection, and an alternative LBPH fallback (VerbaSign).

Results at a Glance

	Custom CNN	LBPH (VerbaSign)
Approach	Deep Learning	Classical ML
Input Size	300x300 RGB	200x200 Grayscale
Model Size	~350 MB	~17 MB
Robustness	High (Illumination, Pose)	Medium
Prediction	WebCam Inference	WebCam Inference

Custom CNN Model (Trained)

• Confidence Threshold: > 0.90
• Loss Function: CrossEntropyLoss
• Optimizer: Adam (lr=0.001)
• Transforms: Resize, ToTensor, Normalize

LBPH Model (Baseline)

• Confidence Threshold: < 70.0
• Algorithm: Local Binary Patterns Histograms
• Setup Time: Very Fast

---

How It Works

• Observation (3-channel RGB): Real-time webcam feed cropped by Haar cascade to 300x300 resolution, normalized, and transformed to PyTorch tensors.
• Action/Prediction (N-dim): Custom CNN outputs raw logits corresponding to num_classes representing different recognized individuals.
• Key preprocessing components: Automated aspect-ratio preserving crop—collect_data.py ensures no distorted faces enter the training set by mathematically centering and padding the cropped bounding box with white pixels.
• Termination/Fallback: If the probability for the predicted class does not exceed the 0.90 confidence threshold, it classifies the face as Unknown.

---

Setup

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

Face_Recognition/
├── config.py            # Configuration paths, parameters
├── collect_data.py      # Face extraction using Haar Cascade
├── train_cnn.py         # Custom CNN PyTorch Training Loop 
├── inference_cnn.py     # Real-time face recognition inference
├── cnn_model.py         # CNN Architecture Definition
├── verify_data.py       # Verification script for corrupted images
└── verbasign.py         # Alternative LBPH approach

---

Usage

# 1. Collect Data (Automatically detects, crops, and centers faces)
python collect_data.py

# 2. Verify dataset integrity
python verify_data.py

# 3. Train the Custom CNN model
python train_cnn.py

# 4. Evaluate & Real-time Inference (CNN)
python inference_cnn.py

# 5. Alternative (LBPH Face Recognizer Collection, Training, & Inference)
python verbasign.py

---

Training Progression

Note: Typical training progression over 2 epochs with Adam optimizer.

Epoch	Batch loss	Train/Val Split
1	~1.5	50%
2	~0.5	50%

---

Key Design Decisions

• Custom PyTorch CNN — flexible CNN architecture using robust tensor transformations (transforms.Normalize) and pooling, achieving high-accuracy spatial feature extraction.
• Higher confidence threshold (0.90) — ensures false positives are strictly classified as "Unknown" for better security.
• Automated aspect-ratio preserving crop — collect_data.py mathematically centers and pads the cropped bounding box with white pixels, so no stretched or distorted faces negatively impact the dataset.
• LBPH Fallback (VerbaSign) — provides an immediate, low-resource deployment alternative without requiring heavy tensor computations.

---

License

MIT — see PyTorch and OpenCV for dependencies.