FDW-YOLOv8: A Lightweight Unmanned Aerial Vehicle Small Target Detection Algorithm Based on Enhanced YOLOv8

The aerial imaging target detection of unmanned aerial vehicles (UAV) stands as a pivotal application, yet extant algorithms for UAV target detection bear significantly heightened computational complexity. Moreover, they grapple with issues of both false positives and negatives in detecting smaller targets. The emergent generation of UAV target detection technology necessitates heightened detection accuracy, minimized computational overhead, accelerated detection speeds, and fortified deployment flexibility. Addressing these requisites, the FDW-YOLOv8 detection model has been proposed. Primarily, our approach involves the enhancement of the PWConv structure and the integration of the FasterNeXt module within the backbone network. By concurrently reducing redundant computations and memory access, we efficaciously extract spatial features, thereby furthering the model's lightweight nature. Subsequently, an additional small target detection branch has been incorporated at the detection layer, coupling with a multi-scale detection method to amplify the model's multi-scale feature fusion capability, effectively refining the network's precision in detecting smaller targets. Furthermore, the use of WIoU Loss optimizes the network's bounding box losses, fostering rapid convergence and augmenting the model's generalization capabilities. Experimental trials conducted on the VisDrone2019 UAV aerial dataset evince that our proposed FDW-YOLOv8 model exhibits a 1.3% enhancement in mean average precision (mAP) over the traditional YOLOv8 model and a 1.8% enhancement in precision (P). Concurrently, post-lightweight modification, the model's parameters have diminished by 18.4%, thereby catering to the deployment demands that necessitate high precision while accommodating low computational resources.