Dy-YOLO: an improved YOLOv8n model for small object detection in VisDrone aerial images
Abstract
Aerial images captured by UAVs suffer from small target sizes, weak features, scattered distributions, and cluttered backgrounds, leading to missed detections and false positives. To address these issues, this paper presents Dy-YOLO (Dynamic Convolution-You Only Look Once), a lightweight enhanced model for detecting small aerial objects built upon YOLOv8n. First, to overcome the limitation of static convolution in extracting weak features, Dynamic Convolution (DyConv) is introduced to reconstruct the C2f module, resulting in a Dy-C2f module that enhances adaptive feature extraction with minimal overhead, thereby reducing information loss during convolution and downsampling. Second, the native detection head is optimized by removing the ineffective 20×20 P5 head and adding a 160×160 P2 head dedicated to small objects. Meanwhile, the neck’s feature fusion path is improved to fully integrate shallow detail features with deep semantic features. Experimental results on the VisDrone-DET2019 dataset show that Dy-YOLO achieves a 3.7 percentage point gain in mAP@50 and a 3.5 percentage point gain in mAP@50:95 relative to the baseline YOLOv8n, using merely 2.45 million parameters. This achieves a favorable accuracy-efficiency trade-off with competitive real-time inference speed, thereby satisfying the deployment requirements of UAV edge devices.