基于改进YOLOv8n的航拍视角小目标检测算法OA

In view of the low detection accuracy caused by small scale,dense and occlusion of the objects in the UAV aerial photography viewpoint,as well as the resource constraints on UAV(unmanned aerial vehicle)equipment,an improved small object detection algorithm for UAV aerial photography viewpoint is proposed based on YOLOv8n.Firstly,the receptive field,and channel and spatial attention mechanisms are integrated into the convolution to design RFCSAMConv(receptive-field channel space attention module convolution)module,so as to improve the model ability of extracting complex features,and then the dilation-wise residual C2f(C2f-DWR)module is also introduced to obtain multi-scale context efficiently and enhance the model capability of recognizing small objects.Secondly,the enhanced adaptive bidirectional multi-scale feature fusion neck network structure is designed,SPDConv(space-to-depth convolution)is used to capture fine-grained features,and the multi-kernel attention MKA(multi-kernel attention)module is designed to enhance multi-scale feature fusion.Finally,Soft-NMS(soft non-maximum suppression)is employed to adjust the overlapping box suppression strategy and improve the detection accuracy of the occluded objects.Experiments on the dataset VisDrone2019 show that,the mean average precision mAP@0.5 and mAP@0.5:0.95 of the improved model are enhanced by 8.7%and 7.6%,respectively,and its parameters are reduced by 31.5%in comparison with those of the benchmark model YOLOv8n,which verify the effectiveness of the improved model in small object detection for UAV aerial photography viewpoint.