An Improved Lightweight YOLOv5 Network for Small Targets Detection

With the continuous development of unmanned systems, Unmanned aerial Vehicles (UAVs) have been paid more and more attention because of their unique advantages in the air. Target detection replaces the traditional human eye detection, which has the advantage of being faster and more accurate. However, the current UAV target detection also has some problems: for example, the target is too small and the detection algorithm is too complex. Based on the existing YOLOv5 network framework, we propose a lightweight model for small target detection. In this paper, a fourth feature extraction channel is added on the basis of the original YOLOv5 three-scale feature detection network. The feature map size in this channel is 160 x 160, which is especially for small target detection. In order to lighten the network and reduce the computation, we introduced Ghost convolution module in YOLOv5 neck network. The improved model is tested on the VisDrone2019 dataset, and the results show that compared with the original YOLOv5 algorithm, the proposed algorithm achieves higher detection accuracy with less computation and model complexity.