A Modified YOLOv4 Deep Learning Network for Vision-Based UAV Recognition

Abstract

The use of drones in various applications has now increased, and their popularity among the general public has increased. As a result, the possibility of their misuse and their unauthorized intrusion into important places such as airports and power plants are increasing, threatening public safety. For this reason, accurate and rapid recognition of their types is very important to prevent their misuse and the security problems caused by unauthorized access to them. Performing this operation in visible images is always associated with challenges, such as the small size of the drone, confusion with birds, the presence of hidden areas, and crowded backgrounds. In this paper, a novel and accurate technique with a change in the YOLOv4 network is presented to recognize four types of drones (multirotors, fixed-wing, helicopters, and VTOLs) and to distinguish them from birds using a set of 26,000 visible images. In this network, more precise and detailed semantic features were extracted by changing the number of convolutional layers. The performance of the basic YOLOv4 network was also evaluated on the same dataset, and the proposed model performed better than the basic network in solving the challenges. Compared to the basic YOLOv4 network, the proposed model provides better performance in solving challenges. Additionally, it can perform automated vision-based recognition with a loss of 0.58 in the training phase and 83% F1-score, 83% accuracy, 83% mean Average Precision (mAP), and 84% Intersection over Union (IoU) in the testing phase. These results represent a slight improvement of 4% in these evaluation criteria over the YOLOv4 basic model.