Transformer-based Bird vs Drone Classification for Anti-drone System: A Strategy and Comparative Study to Reduce False-Positive Alarms

Anti-Drone systems for a long time have been using different architectures to detect, identify, and track drones. Although it has been a great system, it has also seen an increased rate of false positives. Transformer-based architectures like Vision Transformer (ViT) and Swin Transformer (Shifted Window Transformer) offer promising solutions through their self-attention mechanisms. The study aims to evaluate these two state-of-the-art deep learning models' performance, efficiency, and accuracy by employing a comprehensive dataset of diverse bird and drone images. The paper analyzes various metrics such as accuracy, precision, recall, and F1-score through rigorous experimentation, providing insights into their applicability inthe real world. Preliminary findings indicate that the ViT-based framework achieves a training accuracy of 99.16%. In comparison, the Swin Transformer, with its hierarchical feature extraction and shifted window mechanism, excels in handling complex backgrounds has training accuracy of 99.41%. TheViT, as well as the Swin Transformer, demonstrates high precision and recall across all classes, including drones and birds.