A468 — Privacy Shield: Object Recognition & Anonymization

A web-based tool that detects faces and license plates in images, then automatically blurs them for privacy protection.

---

Overview

A468 is a privacy-focused image processing web application. Upload any image, select which objects to detect, and the system will automatically locate and blur them — protecting identities and sensitive information.

Currently supported detection targets:

• Faces — detected using MTCNN + FaceNet (InceptionResNetV1 pretrained on VGGFace2)
• License Plates (in development) — detected using a custom InceptionResNetV2 model trained on annotated plate data

---

Features

Feature	Description
Face Detection & Blurring	Detects all faces in an image using MTCNN and applies elliptical Gaussian blur
License Plate Blurring	Detects and blurs number plates using a custom-trained TF/Keras model (WIP)
User Authentication	Register/login system with bcrypt-hashed passwords stored in SQLite
Image History	Per-user gallery of all previously processed images with timestamps and applied filters
GPU Acceleration	Automatically uses CUDA if available, falls back to CPU

---

Tech Stack

Layer	Technology
Backend	Python, Flask
ML — Face Detection	PyTorch, facenet-pytorch (MTCNN + InceptionResNetV1)
ML — Plate Detection	TensorFlow/Keras, InceptionResNetV2
Image Processing	OpenCV, NumPy, Pillow
Database	SQLite (via Python sqlite3)
Auth	bcrypt
Frontend	HTML, CSS (Jinja2 templates)

---

Project Structure

A468/
├── main.py                  # Flask app — routes, auth, upload, view
├── mtcnn.py                 # FaceModel class — MTCNN detection + blur logic
├── facecnn.py               # Standalone CNN trained on CIFAR-10 (experimental)
├── numplate_model.py        # NumplateModel class — plate detection + blur logic
├── numplate-model.py        # Alternate plate model script
├── numplate-train.py        # Training script for the license plate model
├── requirements.txt         # Python dependencies
├── static/
│   ├── A468.png             # Logo (full)
│   ├── A468_flat.png        # Logo (flat)
│   ├── style.css            # Main stylesheet
│   ├── style_new.css        # Updated stylesheet
│   └── img/                 # Sample/test images
└── templates/
    ├── landing.html         # Public landing page
    ├── login.html           # Login page
    ├── register.html        # Registration page
    ├── main.html            # Upload & result page (authenticated)
    └── view.html            # Image history gallery

---

Installation

Prerequisites

• Python 3.8+
• pip
• (Optional) NVIDIA GPU with CUDA for faster inference

Steps

1. Clone the repository

   git clone https://github.com/XynoxTheDev/A468.git
   cd A468

2. Create and activate a virtual environment (recommended)

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

3. Install dependencies

   pip install -r requirements.txt

   Note: torch, torchvision, and NVIDIA CUDA packages are included in requirements.txt. If you are on CPU only, you can replace the CUDA torch packages with the CPU-only build from pytorch.org.

4. Run the application

   python main.py

5. Open your browser and navigate to http://localhost:5000

---

Usage

1. Register / Log In

Create an account on the registration page. Passwords are hashed with bcrypt before storage.

2. Upload an Image

From the main dashboard, upload any image and select which filters to apply:

• Check Face to detect and blur all faces in the image.
• (License plate blurring is in development)

3. View Results

The processed image is displayed immediately and saved to your personal history.

4. Image History

Navigate to the View page to browse all images you have previously processed, along with timestamps and the filters that were applied.

---

How It Works

Face Detection & Blurring (mtcnn.py)

1. MTCNN detects all face bounding boxes in the image.
2. Each bounding box is expanded (2x) to include surrounding context.
3. A Gaussian blur (kernel=99x99, sigma=15) is applied to the face region.
4. An elliptical mask blends the blurred face naturally back into the original image.

License Plate Detection (numplate_model.py)

1. The image is resized to 224x224 and normalized.
2. A fine-tuned InceptionResNetV2 model predicts bounding box coordinates.
3. Coordinates are denormalized back to original image dimensions.
4. The detected region is blurred using Gaussian blur and composited back.

Training the Plate Model (numplate-train.py)

The number plate model was trained on a custom annotated dataset:

• XML annotations parsed to extract bounding box labels
• InceptionResNetV2 (ImageNet pretrained) used as the base
• Custom regression head: Flatten → Dense(500) → Dense(250) → Dense(4, sigmoid)
• Trained for 180 epochs with MSE loss and Adam optimizer
• Saved as numplate_detection.h5

---

Configuration

Setting	Location	Default
Flask secret key	main.py:14	'a468'
Database file	main.py:26	database.db (auto-created)
Server port	main.py:148	5000
Debug mode	main.py:148	True

Production note: Change the secret key and disable debug mode before deploying.

---

License

This project is licensed under the MIT License. See LICENSE for details.

---

Contributors

• XynoxTheDev
• Anurag3301