Improved YOLOv8 Model for Small Object Detection in Drone Images

Aiming at the problem of small target size, low resolution and complex background in UAV aerial images, an improved UAV aerial small target detection model YOLOv8-IAF based on YOLOv8 is proposed to improve the detection accuracy of small targets in UAV aerial photography. The bottleneck module in C2f is optimised, and the C2f_faster module is proposed to lighten the network and reduce the computation of the network; the Asymptotic Feature Pyramid Network (AFPN) is used to fuse the two low-level features, followed by the gradual incorporation of more high-level features, which makes it more efficient in multi-scale feature fusion and improves the detection of small targets; and the use of the inner-ShapeIoU loss function instead of the original CIoU loss function improves the accuracy of bounding box regression. Experimental validation on the VisDrone2019 dataset shows that YOLOv8-IAF improves mAP50 by 4.2 percentage points, respectively, over the original YOLOv8n algorithm, effectively reducing leakage and misdetection and demonstrating its potential application in small target detection in UAV aerial photography.