🎓 University of Sindh — AI-Based ID Card Detection System

A real-time AI system that analyzes student ID cards via live camera feed and verifies whether the card belongs to the University of Sindh, Jamshoro.

---

📌 Project Overview

This project is an end-to-end deep learning application built for academic exhibition. It uses a fine-tuned MobileNetV2 model trained via transfer learning to perform binary classification on student ID card images — distinguishing University of Sindh cards from all others. The system runs in real-time using a phone camera connected via DroidCam, with results displayed in a clean desktop GUI.

---

✨ Features

• 🔴 Live Camera Detection — Real-time ID card scanning using phone camera (DroidCam)
• 🧠 Transfer Learning Model — MobileNetV2 fine-tuned on custom dataset (97–98% confidence on target cards)
• 🃏 Automatic Card Detection — OpenCV contour detection isolates the card region from the background
• ✅ Instant Verdict — VERIFIED or REJECTED with confidence score displayed live
• 🖥️ Desktop GUI — Built with Tkinter, dark-themed UI with animated confidence bar
• 📦 Lightweight — Runs fully locally, no internet required at inference time

---

🗂️ Project Structure

ID Card Detection System/
├── app.py                          # Main desktop application
├── logo_template.jpeg              # University of Sindh logo crop (for template matching)
├── requirements.txt                # Python dependencies
├── model/
│   └── university_id_classifier.keras   # Trained MobileNetV2 model
└── Cards Data/                     # Training dataset (not included in repo)
    ├── my_uni/                     # University of Sindh card images (~308)
    └── other_uni/                  # Other/CNIC card images (~308)

---

🧠 Model Architecture

Component	Details
Base Model	MobileNetV2 (ImageNet pre-trained)
Training Strategy	Two-phase: frozen base → fine-tune last 30 layers
Input Size	224 × 224 × 3
Output	Binary (sigmoid) — my_uni vs other_uni
Optimizer	Adam (Phase 1: 1e-3, Phase 2: 1e-5)
Regularization	Dropout (0.4, 0.3), BatchNormalization
Trained On	Kaggle (GPU T4)
Val Accuracy	100% on held-out validation set

Data Augmentation Applied

• Rotation (±20°)
• Width/Height shift (10%)
• Zoom (15%)
• Brightness jitter (0.7–1.3)
• Shear (10%)

---

📦 Dataset

Class	Source	Count
my_uni	Real photos of University of Sindh student ID cards (self-collected)	~308
other_uni	Pakistani CNIC cards from Roboflow CNIC Dataset	308

Note: Dataset is not included in this repository due to privacy concerns (real student ID cards).

---

🚀 Getting Started

Prerequisites

• Python 3.10+
• Windows 10/11
• DroidCam app on Android phone (for phone camera)

Installation

# Clone the repository
git clone https://github.com/yourusername/id-card-detection-system.git
cd id-card-detection-system

# Install dependencies
pip install -r requirements.txt

Running the App

1. Open DroidCam app on your phone (keep desktop client closed)
2. Run the application:

python app.py

3. Click ▶ Start Camera in the GUI
4. Hold a University of Sindh ID card in front of the camera

If the camera doesn't start, change CAMERA_INDEX = 1 to CAMERA_INDEX = 0 at the top of app.py

---

⚙️ Configuration

All settings are at the top of app.py:

MODEL_PATH     = "model/university_id_classifier.keras"
LOGO_PATH      = "logo_template.jpeg"
IMG_SIZE       = (224, 224)
CONF_THRESHOLD = 0.90    # Minimum confidence to mark as VERIFIED
CAMERA_INDEX   = 1       # Camera index (1 = DroidCam, 0 = laptop webcam)

---

🏋️ Training Your Own Model

The training notebook is available on Kaggle. To retrain:

1. Organize your dataset:

Cards Data/
├── my_uni/       ← your university card images
└── other_uni/    ← negative class images

2. Upload to Kaggle as a dataset named cards-data
3. Run the training notebook (GPU T4 recommended)
4. Download university_id_classifier.keras from outputs
5. Place in model/ folder

---

🔍 How It Works

Phone Camera (DroidCam)
        ↓
   OpenCV reads frame
        ↓
   Card region detected via contour analysis
        ↓
   Region preprocessed → 224×224, normalized
        ↓
   MobileNetV2 predicts confidence score
        ↓
   Confidence ≥ 90%  →  ✅ VERIFIED
   Confidence < 90%  →  ❌ REJECTED
        ↓
   Result displayed in Tkinter GUI with bounding box on video feed

---

🛠️ Tech Stack

Layer	Technology
Deep Learning	TensorFlow / Keras
Base Model	MobileNetV2 (transfer learning)
Computer Vision	OpenCV
GUI	Tkinter + Pillow
Training Platform	Kaggle (GPU T4)
Camera	DroidCam (Android → Windows)
Language	Python 3.10

---

📋 Requirements

tensorflow==2.19.0
opencv-python==4.11.0.86
Pillow==11.2.1
numpy==2.1.3

---

⚠️ Limitations

• Trained on a small dataset (~308 images per class) — performance may vary on cards from other universities with similar color schemes
• Negative class consists only of Pakistani CNIC cards — a card from another university with a similar blue design might score higher than expected
• Requires good lighting and a reasonably straight card angle for best results
• DroidCam must be running on phone before launching the app

---

🔮 Future Improvements

• Add OCR module (EasyOCR) to verify university name text as a second layer
• Re-introduce logo template matching with a higher-quality logo crop
• Expand negative class with more university card varieties
• Add student face matching against photo on card
• Package as standalone .exe with PyInstaller

---

👨‍💻 Author

Muhammad Haroon Abbas BS Computer Science — University of Sindh, Laar Campus @ Badin Roll No: 2K23/BLCS/38

---

📄 License

This project is for academic/educational purposes only. Not intended for production security use.