🤖 AI vs Human Face Classification

Real-Time Face Detection + ResNet18 Model (PyTorch)

This repository contains two main components:

1. Model Training Script – train_finetune.py
   Fine-tunes a ResNet18 model to classify faces as AI-generated or Human.

2. Real-Time Detection Script – detect_realtime.py
   Uses webcam/screen capture + Haarcascade + PyTorch model to detect faces and classify them live.

---

🔥 Features

✅ Model Training (train_finetune.py)

• Downloads dataset from Google Drive
• Fine-tunes ResNet18
• Early image transforms + dataloaders
• Saves trained model as fine_tuned_ai_vs_human.pth

---

🎥 Real-Time Detection (detect_realtime.py)

• Captures screen region using MSS
• Detects faces using Haarcascade
• Classifies each face as AI or Human
• Confidence display
• Label smoothing using Deque
• Real-time bounding boxes + color coding
  • 🟢 Human
  • 🔴 AI

---

📁 Project Structure

ai-vs-human-detector/ │── detect_realtime.py │── train_finetune.py │── fine_tuned_ai_vs_human.pth │── requirements.txt │── README.md │ └── dataset/ ├── train/ ├── validate/ └── test/

🚀 Installation

1️⃣ Clone the repo

git clone https://github.com/yourusername/ai-vs-human-detector.git

cd ai-vs-human-detector

2️⃣ Install dependencies

pip install -r requirements.txt

🧠 Model Training

Run training script:

python train_finetune.py

Expected Output:

Dataset is downloaded & extracted

ResNet18 is fine-tuned for 5 epochs

Final model saved at: fine_tuned_ai_vs_human.pth

Replace the dataset Google Drive file ID here:

DATASET_URL = "https://drive.google.com/uc?id=YOUR_FILE_ID"

🎥 Real-Time Face Classification

Make sure fine_tuned_ai_vs_human.pth is in the same folder as detect_realtime.py.

Run detection:

python detect_realtime.py

Controls:

Press q → Quit

Real-time window:

Green box = Human

Red box = AI

Smoothed labels for accuracy

⚙️ How It Works

🟩 1. Face Detection

Uses OpenCV Haarcascade to detect faces on screen:

face_cascade = cv2.CascadeClassifier(...)

🟦 2. Face Validation

Removes bad, dark, cropped, invalid faces:

is_valid_face(face_crop)

🟥 3. ResNet18 Inference

Transforms → Tensor → Feed to model:

output = model(input_tensor)

🟨 4. Stable Labeling

Smooths predictions over last 5 frames:

stable_label()

📦 Requirements

torch

torchvision

opencv-python

numpy

mss

pillow

tqdm

gdown

Install all:

pip install -r requirements.txt

🛠 Future Improvements

Add GUI (Streamlit or Flask)

Support webcam detection

Add dataset augmentation

Export model to ONNX for mobile use

📜 License

MIT License – Free to use and modify.

✨ Author

Laksha Karimuthu