Classification of Military Vehicles and Weapons using Deep Learning Architectures

Abstract

In this study, a real-time model that can detect military vehicles, armed and unarmed people was designed with the model trained using deep learning methods. In order to classify objects, 3 different transfer learning architectures such as ResNet50, Mobilenet, EfficientNet were used and performance comparisons were made. A new 6-class hybrid dataset containing 4195 images of military and civilian vehicles, civilian and armed human images was prepared for the architectures. In order not to overfit the models, random zooming, random panning in horizontal direction, random panning in vertical direction, horizontal rotation, rotation with a certain angle and horizontal shift operations were applied to the images in the dataset. In the study, the most successful results in the transfer transfer architectures were obtained with the ResNet50 architecture with an accuracy of 0.9916 and an F1-score value of 0.9911. The data set can be augmented by adding more aerial images of military vehicles. Each class in the dataset can be customized and divided into different classes.