Detection of Small Targets in UAV Images by Using Improved YOLOv8s

The significance of object detection in unmanned aerial vehicles (UAVs) has been drastically increased in various applications, e.g., traffic management, agriculture, military affairs, etc. The you-only-look-once version 8s (YOLOv8s) model is widely adopted in real-time object detection tasks due to its speed and high performance. However, challenges remain when dealing with complex real-world scenarios such as dense scenes with multiple small targets. To address the issue of false or missed detection of small targets in UAV-captured scenes, we improve the YOLOv8s model by incorporating (i) a small target detection layer, (ii) the deformable convolutional networks v2 (DCNv2) module, and (iii) the spatial pyramid pooling fast-large separable kernel attention (SPPF-LSKA) module into it. The new model is called YOLOv8s-ZS and the dataset used is VisDrone2019. The experimental results show that YOLOv8s-ZS achieves an improvement of up to 3.8% in terms of precision, recall, and mean average precision, as compared with some models, e.g., YOLOv8s, BDH-YOLO, and CoT-YOLOv8. These results demonstrate that YOLOv8s-ZS provides a better solution for the problem of UAV small target detection.