Advancing Real-Time Military Aircraft Detection: a Comprehensive Comparative Benchmark of Object Detection Frameworks

Abstract

This paper provides a comprehensive comparative analysis of deep learning and transformer-based models of the state of the art military aircraft detection from aerial imagery. We evaluate the performance of some models like EfficientNetB3, EfficientNetB5, Vision Transformer (ViT), and Swin Transformer on a specialized military aircraft dataset. Our experimental results demonstrate EfficientNetB3 to be 93 % accurate in the classification of military aircraft with computational efficiency, according to the compound scaling principles established by Tan and Le. We also employ RT-DETR for object detection, which is <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$92.7 \% \text{mAP}$</tex> and 90.4 % recall on 81 different types of military aircraft. For comparison, YOLOv8 yielded <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$94 \% \text{mAP}$</tex> and 88.1% recall, and YOLOv7 yielded 90.2% mAP and 82.7 % recall, pushing the real-time detection benchmark established by Redmon et al. further. We use aggressive data augmentation techniques and preprocessing pipelines for enhancing model generalizability.