Brand Detection System

Digital Image Processing Project – Cairo University

A comprehensive logo/brand detection system comparing classical computer vision techniques with modern deep learning approaches, featuring a Flutter mobile application for real-time brand verification.

---

📋 Table of Contents

• Overview
• Supported Brands
• System Architecture
• Prerequisites
• Project Structure
• Installation & Setup
• Running the System
• Evaluation & Testing
• System Workflow
• Results & Analysis

---

🎯 Overview

This project implements two distinct approaches for brand detection:

1. Classical Computer Vision Pipeline

• SIFT (Scale-Invariant Feature Transform) for feature extraction
• Descriptor matching using FLANN-based matcher
• Homography verification for geometric consistency
• Rule-based logo verification for final classification

2. Deep Learning Approach

• Vision Transformer (ViT) architecture
• Pretrained model: Falconsai/brand_identification
• Used for performance comparison and benchmarking

Frontend

• Flutter mobile application for image capture and result visualization
• Real-time communication with backend via USB connection

---

🏷️ Supported Brands

The system currently detects three major sports brands:

• Nike
• Adidas
• Puma

Output Format

✓ REAL → BRAND NAME (Nike/Adidas/Puma)
✗ FAKE → UNKNOWN BRAND

---

🏗️ System Architecture

┌─────────────────┐
│  Flutter App    │
│  (Mobile)       │
└────────┬────────┘
         │ USB Connection
         ▼
┌─────────────────┐
│  Python Backend │
│  (Flask Server) │
└────────┬────────┘
         │
    ┌────┴────┐
    ▼         ▼
┌────────┐ ┌────────┐
│   CV   │ │   DL   │
│ SIFT   │ │  ViT   │
└────────┘ └────────┘

---

⚠️ Prerequisites

Required Hardware & Software

Component	Requirement	Purpose
Python	3.9 or higher	Backend processing
Android Device	Physical phone	Mobile app deployment
USB Cable	Type-C	Device-laptop connection
Flutter SDK	Latest stable	Mobile app development
Connection	USB only	Socket communication

Critical Requirements

⚠️ IMPORTANT: The system will NOT work over Wi-Fi. USB connection is mandatory for stable socket communication.

Before Running:

1. ✅ Backend server must be running
2. ✅ Mobile app must be installed
3. ✅ Phone must be connected via USB
4. ✅ Run app in release mode for optimal performance

---

📁 Project Structure

Brand-Detection/
├── src/                          # Python backend server
│   ├── main.py                   # Flask server entry point
│   ├── test_brand_detection.py   # Classical CV testing
│   └── run_brand_dnn.py          # Deep learning testing
│
├── Logos/                        # Reference brand logos
│   ├── nike.png
│   ├── adidas.png
│   └── puma.png
│
├── Experiment Results/           # Test dataset
│   ├── real/                     # Authentic brand images
│   │   ├── nike/
│   │   ├── adidas/
│   │   └── puma/
│   └── fake/                     # Non-brand/counterfeit images
│
└── frontend/                     # Flutter mobile application
    └── brand_detection/
        └── lib/                  # Dart source code

---

🔧 Installation & Setup

1. Backend Setup

# Navigate to backend directory
cd Brand-Detection/src

# Install dependencies
pip install -r requirements.txt

# Verify installation
python --version  # Should be 3.9+

2. Frontend Setup

# Navigate to Flutter project
cd Brand-Detection/frontend/brand_detection

# Get Flutter dependencies
flutter pub get

# Verify Flutter installation
flutter doctor

---

🚀 Running the System

Step 1: Start the Backend Server

cd Brand-Detection/src
python main.py

Expected Output:

* Running on http://127.0.0.1:5555
* Server started successfully

Step 2: Launch the Mobile App

install mobile app from github realese

Why Release Mode?

• ⚡ Better performance
• 🔄 Stable socket communication
• 📸 Accurate real-time image transfer
• 🐛 Fewer debugging overheads

Step 3: Connect and Test

1. Connect Android device via USB
2. Enable USB debugging on phone
3. Launch app and capture/select brand image
4. View real-time detection results

---

🧪 Evaluation & Testing

Classical Computer Vision Evaluation

Test the SIFT-based approach on the experimental dataset:

cd Brand-Detection/src
python test_brand_detection.py

What This Does:

• Processes all images in Experiment Results/
• Applies SIFT feature extraction and matching
• Verifies homography for geometric consistency
• Generates accuracy metrics and detailed logs

Evaluation Criteria:

Image Type	Correct Prediction
Real Photos	Output matches folder name (nike/adidas/puma)
Fake Photos	Output is NOT one of the target brands

Deep Learning Model Evaluation

Compare results with the pretrained Vision Transformer:

cd Brand-Detection/src
python run_brand_dnn.py

What This Does:

• Uses Falconsai/brand_identification model
• Classifies all test images
• Validates predictions against ground truth
• Produces comparative accuracy metrics

Performance Metrics

Both approaches are evaluated using:

• ✅ Accuracy (overall correct predictions)
• 📊 Precision & Recall (per-brand performance)
• 🎯 Success cases (correctly classified images)
• ❌ Failure cases (misclassifications with analysis)
• 📝 Detailed logging (for debugging and improvement)

---

🔄 System Workflow

1. User captures/selects image in Flutter app
          ↓
2. Image sent to backend via USB socket
          ↓
3. Backend processes with CV/DL pipeline
          ↓
4. Brand classification performed
          ↓
5. Result returned to mobile app
          ↓
6. User views: REAL → BRAND or FAKE → UNKNOWN

---

📊 Results & Analysis

Approach Comparison

Aspect	Classical CV (SIFT)	Deep Learning (ViT)
Speed	Fast (milliseconds)	Moderate (model inference)
Accuracy	High for clear logos	Very high overall
Robustness	Sensitive to occlusion	Handles variations well
Resource	Low computational cost	Requires more memory/GPU
Interpretability	High (feature matching)	Low (black box)

Strengths & Weaknesses

Classical CV Strengths:

• ⚡ Fast inference
• 🔍 Explainable feature matching
• 💾 Low resource requirements

Classical CV Weaknesses:

• ❌ Struggles with heavy rotation/distortion
• 🌫️ Poor performance in low contrast
• 📐 Requires good logo visibility

Deep Learning Strengths:

• 🎯 High accuracy on diverse images
• 🔄 Handles variations robustly
• 📸 Works with partial logos

Deep Learning Weaknesses:

• ⏱️ Slower inference
• 🖥️ Requires more computational resources
• 🔒 Less interpretable decisions

---

🎓 Conclusion

This project successfully demonstrates:

• End-to-end brand detection system from image capture to classification
• Comparative analysis of traditional vs. modern computer vision techniques
• Real-world mobile application with practical USB-based communication
• Comprehensive evaluation framework for both approaches

The system highlights that while classical computer vision techniques offer speed and interpretability, deep learning models provide superior accuracy and robustness in varied real-world conditions.

---