UAV Small Target Detection Based on Improved YOLOv8 Algorithm

Unmanned Aerial Vehicles (UAVs) find extensive use across multiple industries, such as agriculture, aerial photography, rescue, and surveillance. But it can be difficult to identify small targets in high-resolution UAV data, which frequently leads to poor accuracy and high miss rates. This work proposes an enhanced YOLOv8 model to overcome this problem. In order to improve small target recognition capabilities, this model incorporates a DCN module into its backbone network and dynamically modifies the receptive field. Moreover, the neck area of the model incorporates CBAM, which makes it possible for salient image regions to be automatically noticed, improving detection performance. Finally, we substitute WIoU for the conventional loss function, improving the accuracy of the bounding box regression for tiny targets. Experiments on the VisDrone2019 dataset demonstrate that our model outperforms the YOLOv8s baseline, increases [email protected] by 3.8% and mAP@[.5:.95] by 1.7%, and outperforms other widely used object detection methods.