🎥 Real-Time Face & Blur Detection System

A real-time computer vision system that detects human faces, evaluates blur quality, and streams live annotated video to a web browser using Flask.

This project focuses on classical computer vision + real-time systems engineering, built as part of a pre-deep-learning CV foundation series.

---

🔍 Demo

---

🚀 Features

• Real-time webcam capture using OpenCV
• Face detection using Haar Cascades
• Blur detection using Laplacian variance
• CLEAR / BLURRY classification with thresholding
• Live MJPEG video streaming via Flask
• Event-based frame saving (only when blur detected)
• Detection logs with timestamps
• Stable resource management (camera lifecycle handling)

---

🧠 System Architecture

graph TD;
A["Webcam"] --> B["OpenCV Frame Capture"];
B --> C["Grayscale Conversion"];
C --> D["Face Detection (Haar Cascade)"];
D --> E["Blur Detection (Laplacian Variance)"];
E --> F["Annotation (Boxes + Labels)"];
F --> G["MJPEG Stream (Flask)"];
G --> H["Browser Display"];

Loading

---

Key frames are saved and detection events are logged in parallel without interrupting the video stream.

---

🛠 Tech Stack

• Python
• OpenCV – image processing & face detection
• Flask – backend web server
• NumPy – numerical operations
• MJPEG Streaming – real-time video delivery

---

▶️ How to Run

1. Clone the repository

git clone <repo-url>
cd project-opencv

2. Create virtual environment

python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

3. Install dependencies

pip install -r requirements.txt

4. Start the Flask app

python src/flask_app.py

5. Open in browser

http://127.0.0.1:5000

📁 Project Structure

project-opencv/
│
├── src/
│   ├── run_detection.py        # Local OpenCV pipeline
│   ├── flask_app.py            # Live browser streaming
│
├── assets/
│   └── haarcascade_frontalface_default.xml
│
├── saved_frames/
│   └── blur_YYYYMMDD_HHMMSS.jpg
│
├── logs/
│   └── detections.log
│
├── requirements.txt
└── README.md

👉 Key Engineering Decisions

• Event-based frame saving instead of saving every frame (prevents I/O overload)
• Generator-based MJPEG streaming for efficient real-time delivery
• Camera ownership inside generator to avoid resource conflicts
• Cooldown logic to prevent repeated frame captures
• Defensive programming to avoid silent stream crashes

👉 Learning Outcomes

• Understanding classical CV before deep learning
• Building stable real-time video pipelines
• Integrating OpenCV with web backends
• Debugging generator lifecycle & hardware resources
• Designing CV systems with observability (logs + events)

---

Project Closure

This project marks the completion of a pre-deep-learning computer vision foundation.

By intentionally using classical techniques (Haar cascades, Laplacian variance) and focusing on real-time system behavior, this project builds the intuition required to understand why modern CV models (YOLO, ViTs) were needed — not just how to use them.

Rather than optimizing for accuracy, the emphasis was on:

• pipeline design
• system stability
• resource management
• real-world failure modes

This foundation makes the transition to modern deep learning–based vision systems significantly more meaningful and less opaque.

---

🚀 Next Direction

• Upgrade face detection to YOLO / DNN-based models
• Explore Transformer-based vision models (ViT)
• Build multi-modal systems combining CV + NLP

This repository is intentionally closed in scope to preserve conceptual clarity.

---