Comparing Convolutional Neural Network(CNN) models for machine learning-based drone and bird classification of anti-drone system

Abstract

As drones become more advanced and commercialized, crimes using drones are also on rise. For this reason, development of anti-drone systems is increasing. In this paper, CNN model is examined that is suitable for visible camera-based drone identification. The CNN models used for the validation are Alexnet, GoLeNet, Inception-v3 Vg16, Resnet-18, Resnet-50 and Squezezenet. These seven models have already been validated in the ImageNet Large Scale Visual Recognition Competition (ILSVRC). In ILSVRC, 1000 labels are classified, but in this study limits them to three drones, birds and backgrounds. Therefore, it is necessary to verify whether the three labels are the same as the ILSVRC result. In order to verify this, CNN models are learned and tested in the same environment. The experimental results show that the performance of Alexnet, Resnet and Squeeznet is relatively better then the others, unlike the performance of CNN known through ILSVRC. his result shows that a shallow network with a simple structure is more reasonable when the number of labels is small. Based on these results, the further work is to develop a neural network optimized for Drone identification.