基于YOLOv8s的轻量化航拍图像小目标检测算法

邬开俊,郑云琦,魏鼎,袁海翔

YOLOv8s based lightweight algorithm for small object detection in aerial imagery

Kaijun WU,Yunqi ZHENG,Ding WEI,Haixiang YUAN

表 4 VisDrone测试集上的对比实验结果

Tab.4 Comparative tests results on VisDrone test set

模型 Params/106 AP/% AP50/% AP75/% APS/% APM/% APL/% FLOPs/109 RetinaNet[3] 21.37 8.0 15.5 7.6 2.1 13.2 23.7 15.9 FasterR-CNN[4] 41.72 12.8 23.9 12.6 5.2 21.1 29.7 26.8 YOLOXm[24] 12.42 13.9 24.5 14.1 6.0 22.0 23.9 18.1 TOOD[25] 15.64 16.9 28.2 16.8 7.9 24.3 43.1 43.8 VFNet[26] 9.84 17.6 30.3 17.9 8.0 26.7 36.4 10.2 YOLOX-Tiny 5.04 14.8 27.8 14.9 7.6 22.1 27.8 7.6 YOLOv8m 25.85 19.0 33.2 19.2 9.0 29.4 41.7 79.1 YOLOv10m 15.32 19.5 34.5 19.6 9.7 30.0 41.4 58.9 YOLO11s[22] 9.42 17.6 31.3 17.8 8.0 27.2 36.4 21.3 YOLO12s[23] 9.23 17.6 31.2 17.8 8.1 27.4 35.6 21.2 RetinaNet-R50-FPN 36.52 16.4 27.6 14.8 6.0 27.4 42.7 210 GFL[27] 32.28 19.1 32.1 19.2 9.1 30.0 40.9 206 本文方法 9.33 20.9 36.9 21.1 11.0 31.4 42.8 38.3