基于KFSI-YOLOv5的无人机航拍图像遮挡小目标检测算法OA

A KFSI-YOLOv5 algorithm is proposed to address the problem of small and heavily occluded targets in aerial images.Firstly,K-Means++clustering is used to generate optimal anchor boxes suitable for aerial targets.Secondly,the Fp-SEAM attention mechanism is proposed and fused with the FPN P2 layer to improve the detection accuracy of occluded small targets.Then a dynamic label assignment strategy is adopted,which can assign sub-optimal labels for severely occluded targets.Finally,an α-Inner IoU loss function is proposed to make bounding box regression more accurate.Experiments on the VisDrone,Carpk and WiderPerson datasets verify that the proposed algorithm can effectively accomplish the task of occluded small target detection in aerial images.