Improved multi-scale small target detection by UAV

For UAVs, it is usually challenging to carry out target detection tasks. It is particularly difficult to enhance the detection accuracy of densely distributed small targets. Based on YOLOV8, this paper proposes an improved small target detection model. Firstly, a small target detection head is added to YOLOV8 to improve the model’s detection accuracy. Secondly, a weighted down-sampling fusion module (WDF) is designed to augment the feature extracting ability for small targets. In addition, a multi-scale lightweight detection head is developed, which reduces the amount of calculation. Finally, a new dynamic loss function WSIOU is proposed to further improve the detection accuracy. Experimental results on a benchmark dataset for UAVs’ small target detection (VisDrone2021) show that compared with the original YOLOV8 model, the proposed model has significantly reduced the number of parameters by 43.2% and greatly improved mAP50 by 6.4%, which indicates that the improved model is more suitable for UAV small target detection tasks. Furthermore, extensive experimental studies also show that the proposed model has huge advantages in UAVs’ small target detection over prevalent models. In addition, the generalization performance of the proposed model is verified further on a large scale small target dataset UAVDT.