基于改进yolov5的光学遥感图像风车目标检测研究  

Research on Windmill Detection in Optical Remote Sensing Image Based on Improved yolov5

作　　者：薛继伟[1] 吕福娟 XUE Jiwei;LYU Fujuan(School of Computer&Information Technology,Northeast Petroleum University,Daqing 163318)

机构地区：[1]东北石油大学计算机与信息技术学院,大庆163318

出　　处：《计算机与数字工程》2025年第1期245-249,268,共6页Computer & Digital Engineering

摘　　要：遥感图像风车的目标检测有广泛的应用前景,与自然场景图像相比,卫星遥感图像风车存在尺度差异大、目标较小、背景复杂等特点。针对卫星图像检测中存在的干扰误检较多、复杂背景下准确率较低以及检测置信度较低的问题,提出一种基于yolov5的加权双向特征金字塔网络(BiFPN),引入可学习权值来学习不同输入特征,实现了自上而下与自下而上的深浅层特征双向融合,增强不同网络层之间特征信息的传递,在一定程度上提高了多尺度以及受地表覆盖影响下的风车目标识别率,并在同种情况下提高了目标检测的置信度,最终在遥感风车数据集上获得了93.6%的mAP,以及95.3%的精准率和93.0%的召回率。相比于原yolov5网络结构,mAP提升了1.6%,证明了网络改进对复杂背景下遥感图像目标检测的有效性。The windmill detection in optical remote sensing image has a wide range of application prospects.Compared with natural scene images,satellite remote sensing image has the characteristics of large scale differences,smaller targets and complex backgrounds.Aiming at the problems of many interference false detection,low accuracy in complex backgrounds and low detection confidence in satellite image,a weighted bidirectional feature pyramid network(BiFPN)based on yolov5 is proposed,which intro⁃duces learnable weights to learn different input features.It realizes the two-way fusion of top-down and bottom-up deep and shallow features,enhances the transfer of feature information between different network layers,improves the target recognition rate in multi-scale and complex backgrounds,and improves the confidence of object detection.Finally,93.6%mAP,95.3%precision and 93.0%recall are obtained on the remote sensing windmill dataset.Compared with the original yolov5 network structure,the mAP is improved by 1.6%,which proves the effectiveness of the network improvement for target detection in the complex background of re⁃mote sensing images.

关 键 词：光学遥感图像 风车 目标检测 BiFPN 特征融合 

分 类 号：TN911.73[电子电信—通信与信息系统]