Improving the Accuracy of UAV Detection Through Combination of Different Convolution Units

The YOLO (You Only Look Once) series is widely used in various detection fields. This article applies the latest released YOLOv9 and YOLOv10 models to surface-to-air/air-to-air UAV (Unmanned Aerial Vehicle) target detection tasks, assisting anti-UAV tasks and other airborne threat detection tasks. We propose selecting different convolution modules in network layers corresponding to various scales. Specifically, this article uses the DroneDetection dataset, expands the training set using data augmentation techniques, and trains multiple network models using transfer learning techniques. The experimental results show that compared with the most widely used YOLOv3 and YOLOv5 models, YOLOv9 and YOLOv5 models have significant advantages in speed due to their simplification of the network backbone, but have a significant decrease in detection accuracy represented by mAP 50-95 values. In addition, the combination of different convolution modules can improve the accuracy of UAV small target detection.