• Traditional Methods Document Link - https://docs.google.com/document/d/1wXyzUumSO5fcJrvAGRYmBu-czYcWOgaXK156vznpnt8/edit?usp=sharing

1 – Online Face Detection and Recognition System (Traditional Methods)

📌 Overview

This project implements a real-time face detection and recognition system using classical computer vision techniques instead of deep learning. We combined Viola–Jones and HOG for face detection and implemented three recognition models:

1. Eigenfaces + SVM (PCA-based)
2. Fisherfaces (PCA + LDA)
3. LBPH (Local Binary Pattern Histogram)

A Next.js frontend and FastAPI backend were developed to integrate all models into a GUI, enabling real-time webcam-based detection and recognition.

---

⚙ Features

• Two-stage detection: Viola–Jones Haar Cascade + HOG-based detector for robustness.
• Multiple recognition models for performance comparison.
• Custom dataset of 8 members (~70 images per person), with preprocessing and augmentation.
• Real-time GUI for live video-based recognition.
• Evaluation with confusion matrices and classification reports.

---

📂 Models

Model 1 – Eigenfaces + SVM

• PCA to extract Eigenfaces, retaining 95% variance.
• RBF-kernel SVM for classification.
• Accuracy: 93.6% on test set.

Model 2 – Fisherfaces

• PCA for dimensionality reduction, followed by LDA for class separability.
• Accuracy: ~72%.

Model 3 – LBPH

• Local texture-based features using 8x8 grid, radius=1, 8 neighbors.
• Robust to lighting changes.
• Accuracy: ~100% on test set with confidence threshold.

---

🗂 Dataset Preparation

• Videos captured for each member → frames extracted → faces detected.
• Augmentation: flipping, brightness/contrast adjustment, rotation, translation, noise.
• Resizing and histogram equalization for uniformity.

---

🖥 GUI Implementation

• Frontend: Next.js interface for model selection and live camera streaming.
• Backend: FastAPI server integrating all three recognition models.
• Real-time predictions displayed with bounding boxes and labels.

---

📊 Results Summary

Model	Accuracy	Strengths	Weaknesses
Eigenfaces+SVM	93.6%	Strong generalization, compact features	Sensitive to lighting
Fisherfaces	~72%	Better class separation than PCA	Lower accuracy
LBPH	~100%	Very robust to lighting and textures	May struggle with large pose changes

---

• Deep Learning Method Document Link - https://docs.google.com/document/d/1a4mZIMOEuXfyujIfHLMPy6wU-lK1ekDzj-48aa1vn_M/edit?usp=sharing

This project develops two high-accuracy face detection and recognition systems using deep learning, integrated into a Next.js frontend and FastAPI backend for real-time performance.

System 1: MTCNN + FaceNet + FaceClassifier

• Face Detection: MTCNN extracts and aligns faces.
• Embedding Generation: InceptionResNetV1 (FaceNet) creates 512-D feature vectors.
• Classification: Custom classifier predicts identities from embeddings.
• Training: Heavy data augmentation for robustness; 92.41% validation accuracy.

System 2: DeepFace Framework

• Enrollment: RetinaFace detector + ArcFace model to store averaged embeddings.
• Real-time Recognition: YuNet detector for speed; cosine similarity for matching.
• Performance: Flexible, modular, and easily updatable pipeline.

📂 Dataset

• 320 images (40 per person) across 8 classes.
• Captured from multiple angles and stored in PNG format.
• 75:25 train-validation split.

⚙️ Key Features

• Supports both traditional (MTCNN + FaceNet) and DeepFace-based pipelines.
• Frontend GUI allows mode selection, uploads, and displays recognition results.
• Backend optimised for low latency and scalable concurrent requests.

📊 Results

• System 1: 92.41% validation accuracy, strong robustness to augmentation noise.
• System 2: Fast and accurate real-time recognition with modular detector switching.

🚀 Future Work

• Expand dataset diversity.
• Explore alternative face detection methods.
• Fine-tune models for higher accuracy.
• Deploy to mobile/edge devices with lightweight formats like CoreML.