Drone vs Non-Drone Image Classification

Problem

Binary classification of aerial images into drone and non-drone (bird) for automated aerial threat detection.

Dataset

• Total images: 4,106
• Bird: 1,607
• Drone: 2,499
• Moderate class imbalance (~1.55×)
• Random 80/20 train–validation split

Model

• Architecture: ResNet-18
• Pretrained on ImageNet
• Final fully connected layer replaced for binary classification
• Transfer learning used due to limited dataset size

Training Setup

• Input resolution: 224×224
• Optimizer: Adam
• Learning rate: 1e-4
• Loss: Class-weighted CrossEntropyLoss
• Data augmentation: Random horizontal flip
• Training epochs: 5
• Platform: Google Colab (GPU)

Results

• Best validation accuracy: 99.39%

• Confusion matrix:

  • Bird → Drone: 3
  • Drone → Bird: 7

Ablation Study

Frozen Backbone vs Fine-Tuning

• Frozen backbone accuracy: 94.40%
• Fine-tuned accuracy: 99.39%

This confirms that end-to-end fine-tuning significantly improves performance over feature extraction alone.

Limitations

• Dataset likely contains background bias
• No separate test set provided
• Performance may degrade in real-world sky conditions

Limitations Dataset likely contains background bias No separate test set provided Performance may degrade in real-world sky conditions