Human vs AI-generated image classification using Deep Learning and Ensemble approach

This repository contains an advanced deep learning framework for detecting AI-generated images. The system leverages an ensemble approach combining multiple vision models with metadata analysis to achieve high accuracy classification between AI-generated and human-created images.

The core of the system is the AdvancedCLIPDetector which employs:

• Multiple CLIP variants (ViT-L-14, ViT-H-14, ViT-B-16) with different pretraining strategies
• EfficientNetV2-L
• ConvNext Base
• Metadata feature extraction
• Attention-based feature weighting
• Multi-layer classifier with dropout and residual connections

• Advanced Image Processing:

  • Sophisticated image augmentation using RandAugment, ColorJitter, and RandomErasing
  • Test Time Augmentation (TTA) during inference
  • Comprehensive metadata extraction

• Robust Training Pipeline:

  • K-fold cross-validation (default: 5 folds)
  • Focal Loss for class imbalance handling
  • Mixed precision training
  • Gradient clipping
  • Cosine learning rate scheduling with warmup
  • Component-specific learning rates

• State-of-the-Art Inference:

  • Ensemble predictions across folds
  • Soft voting mechanism for final classification
  • TTA during validation and testing

• Gradient analysis using Sobel operators
• JPEG compression artifact detection
• Statistical feature extraction
• Attention mechanisms for feature importance weighting

# Clone the repository
git clone https://github.com/yourusername/ai-image-detection.git
cd ai-image-detection

# Install dependencies
pip install -r requirements.txt

from detector import AdvancedCLIPDetector

# Initialize the detector
detector = AdvancedCLIPDetector(
    pretrained=True,
    num_folds=5
)

# Detect if an image is AI-generated
result = detector.predict("path/to/image.jpg")
probability = result['ai_probability']
is_ai = result['is_ai']  # Boolean classification

print(f"AI-generated probability: {probability:.2f}")
print(f"Classification: {'AI-generated' if is_ai else 'Human-created'}")

from detector import AdvancedCLIPDetector, DataLoader

# Prepare your dataset
train_loader = DataLoader("path/to/training/data", batch_size=32)
val_loader = DataLoader("path/to/validation/data", batch_size=32)

# Initialize and train the model
model = AdvancedCLIPDetector(pretrained=False)
model.train(
    train_loader=train_loader,
    val_loader=val_loader,
    epochs=30,
    learning_rate=1e-4,
    weight_decay=1e-5
)

# Save the trained model
model.save("path/to/save/model")

• Python 3.8+
• PyTorch 1.10+
• torchvision
• timm
• open_clip_torch
• PIL
• numpy
• pandas
• scikit-learn
• albumentations

MIT License

Contributions are welcome! Please feel free to submit a Pull Request.