摘要
Abstract
In view of the low detection accuracy and large error in detecting small-scaled vehicles in the existing UAV aerial image object detection algorithms,a UAV vehicle detection algorithm based on the improved YOLOv8 is proposed,and it is named Improve-YOLOv8.Firstly,a deformable convolutional module DCNv2 is introduced into the C2f convolutional layer of the backbone network,so as to improve the ability of the backbone network to adapt to irregular space structure and enhance the ability of the model to detecting the occluded and overlapped small objects.Secondly,an SPPF-LSKA module with long-range dependence and adaptive ability is proposed on the basis of the idea of Large Separable Kernel Attention,which effectively reduces the background interference on aerial image detection.And then,by introducing DyHead detection head,the three attention mechanisms of scale,space and task are integrated to improve the model detection performance.Finally,WIoUv3 is used as a bounding box regression loss,and a wise gradient allocation strategy is adopted to improve the positioning ability of the model.The experimental results show that in comparison with the benchmark model,the accuracy rate,recall rate and average precision(AP)of the Improve-YOLOv8 are improved by 5.1%,6.1%and 5.1%on the Mapsai dataset,respectively,showing good detection performance and practical application potential.关键词
无人机航拍图像/小目标/YOLOv8/目标检测/可变形卷积/注意力机制Key words
UAV aerial image/small object/YOLOv8/object detection/deformable convolution/attention mechanism分类
电子信息工程
