CNN Based Image Classification of Malicious UAVs

Abstract

Unmanned Aerial Vehicles (UAVs) or drones have found a wide range of useful applications in society over the past few years, but there has also been a growth in the use of UAVs for malicious purposes. One way to manage this issue is to allow reporting of malicious UAVs (e.g., through a smartphone application) with the report including a photo of the UAV. It would be useful to able to automatically identify the type of UAV within the image in terms of the manufacturer and specific product identification using a trained image classification model. In this paper, we discuss the collection of images for three popular UAVs at different elevations and different distances from the observer, and using different camera zoom levels. We then train 4 image classification models based upon Convolutional Neural Networks (CNNs) using this UAV image dataset and the concept of transfer learning from the well-known ImageNet database. The trained models can classify the type of UAV contained in unseen test images with up to approximately 81% accuracy (for the Resnet-18 model), even though 2 of the UAVs represented in the UAV image dataset are visually similar, and the fact that the UAV image dataset contains images of UAVs that are a significant distance from the observer. This provides a motivation to expand the study in the future to include more UAV types and other usage scenarios (e.g., UAVs carrying loads).