Small target detection of Unmanned Aerial Vehicles based on GDWNet in the digital economy

In the era of the rapid development of the digital economy, unmanned aerial vehicle (UAV) aerial imagery offers the advantage of a high-altitude perspective, providing a broader field of view. However, it also presents challenges such as complex backgrounds and small target sizes. To address these issues, a novel UAV aerial image object detection model, GDWNet, is proposed. In the feature extraction stage, GDWNet adopts a multi-branch extraction and merging structure based on the Generalized Efficient Layer Aggregation Network to enhance feature representation and improve gradient flow within the network. During the feature fusion stage, DySample is employed to generate content-aware offsets through learning. This approach effectively breaks the fixed interpolation rules of traditional upsampling methods, enabling more dynamic, flexible, and semantically rich upsampling of the input feature maps, thereby improving the quality of feature integration. The introduction of WIoU_v3 allows the network to focus more on small and medium-sized anchor boxes, addressing the imbalance between ground truth and predicted boxes, and enhancing both the convergence speed and accuracy of the network. When applied to the UAV aerial imagery VisDrone dataset, the proposed model achieved a mean Average Precision (mAP) of 35.8%, representing a 3.1% improvement over the conventional YOLOv8s algorithm.

Text 122515.pdf - Published Version Restricted to Repository staff only Download (1MB) Official URL or Download Paper: https://link.springer.com/article/10.1007/s40032-0...

Actions (login required)

View Item