Abstract: Aiming at the problems of oversized models, slow detection speeds, and high complexity in existing unmanned aerial vehicle (UAN)  target detection algorithms, we  proposed an improved lightweight UAN  detection algorithm YOLO-LDD based on YOLOv5n.  Firstly, on the basis of YOLOv5n, a diversified branch  module DBB and C3 module were introduced to  fuse and  reconstruct into  C3_DBB module, enhancing the representational capacity of individual convolutions. Secondly, a reparameterized structure convolution RepConv was introduced into the neck network to improve detection speed. Finally, the model was compressed by using the layer-adaptive magnitude-based pruning (LAMP) method to reduce the number of parameters. Experimental results show  that the proposed algorithm can maintain excellent detection performance while reducing computational and storage demands, and improve  efficiency and inference speed of the model. The  average accuracy reaches  96.7%, the  parameter count is reduced by 73% compared to YOLOv5n, the computational load is reduced by 60%,  and a detection speed is increased by  1.6 times.

Key words: unmanned aerial vehicle, target detection, YOLOv5n algorithm, lightweight, deep learning