多尺度特征融合的雾霾环境下车辆检测①王忠美1,薛子豪1,伍宣衡1,郑良21(湖南工业大学轨道交通学院,株洲412007)2(中国电子科技集团公司第十五研究所,北京100089)通信作者:王忠美,E-mail:ldwangzm2008@163.com摘要:针对雾霾环境下车辆检测准确率低、漏检严重的问题,提出一种多尺度特征融合的雾霾环境下车辆检测算法.首先利用条件生成对抗网络对雾霾图像进行去雾预处理,然后针对雾霾环境下目标特征不明显的特点,提出多尺度特征融合模块,在YOLOv3的基础上,从主干网络提取特征时增加一条浅层分支和深层特征进行上采样拼接融合,得到尺度为104×104的特征图,用于增强浅层的语义信息.并采用CBAM注意力机制引导下的特征增强策略,保证上下文信息的完整性,以提高检测的精度,最后将去雾后图片送入改进后的YOLOv3网络进行检测.实验结果表明,相较于原始网络,该算法在RTTS数据集上的检测结果更加优秀,模型可以达到81%的平均精度和67.52%的召回率,能够更加精确的定位到车辆.关键词:图像处理;雾霾环境;YOLOv3;注意力机制;特征融合;目标检测引用格式:王忠美,薛子豪,伍宣衡,郑良.多尺度特征融合的雾霾环境下车辆检测.计算机系统应用,2023,32(2):217–225.http://www.c-s-a.org.cn/1003-3254/8957.htmlMulti-scaleFeatureFusionforVehicleDetectioninHazeEnvironmentWANGZhong-Mei1,XUEZi-Hao1,WUXuan-Heng1,ZHENGLiang21(CollegeofRailwayTransportation,HunanUniversityofTechnology,Zhuzhou412007,China)2(The15thResearchInstitute,ChinaElectronicsTechnologyGroupCorporation,Beijing100089,China)Abstract:Givenlowvehicledetectionaccuracyandseriousmissdetectioninahazeenvironment,avehicledetectionalgorithmwithmulti-scalefeaturefusioninahazeenvironmentisproposed.Firstly,theconditionalgenerationandadversarialnetworkisemployedtopreprocessthehazeimages.Then,astheobjectfeatureisnotobviousinahazeenvironment,amulti-scalefeaturefusionmoduleisputforward.OnthebasisofYOLOv3,ashallowbranchisaddedforupsamplingsplicingandfusingwithdeeplayerfeaturesduringextractingfeaturesfrombackbonenetworks.Asaresult,thefeaturemapwiththescaleof104×104isobtained,whichisadoptedtoenhancetheshallowsemanticinformation.ThefeatureenhancementstrategyguidedbytheCBAMattentionmechanismisutilizedtoensuretheintegrityofcontextinformationan...
