电子测量技术ELECTRONICMEASUREMENTTECHNOLOGY第45卷第23期2022年12月DOI:10.19651/j.cnki.emt.2209953自适应特征融合的轻量级交通标志检测方法梁秀满邵彭娟刘振东赵恒斌(华北理工大学电气工程学院唐山063210)摘要:针对目前交通标志检测方法中网络计算量大、检测效果差的问题,提出一种嵌入坐标注意力机制的轻量级交通标志检测方法。首先在MobileNetv2的残差块中嵌入坐标注意力机制CA(channelattention)模块以保留通道注意力中的坐标信息;其次利用改进的MobileNetv2对YOLOv4主干网络做轻量化处理,并且在PANet中采用深度可分离卷积块降低计算量;然后使用ASFF自适应特征融合改进PANet结构来均衡不同特征层的不一致性,最后在特征融合模块加入注意力以增加目标信息的权重;并由K-means++算法产生新的先验框聚类中心。实验表明,权重文件由136M降至54.5M削减了60%,网络体积削减了80%,精度达到96.84%,与YOLOv4网络相比仅损失了0.46%的精度。关键词:轻量级网络;注意力机制;聚类算法;自适应特征融合中图分类号:TP391.4文献标识码:A国家标准学科分类代码:520.6AlightweighttrafficsigndetectionmethodbasedonadaptivefeaturefusionLiangXiumanShaoPengjuanLiuZhendongZhaoHengbin(SchoolofElectricalEngineering,NorthChinaUniversityofScienceandTechnology,Tangshan063210,China)Abstract:Aimingattheproblemsoflargeamountofnetworkcomputationandpoordetectioneffectinthecurrenttrafficsigndetectionmethod,alightweighttrafficsigndetectionmethodwithembeddedcoordinateattentionmechanismisproposed.First,thecoordinateattentionmechanismCAmoduleisembeddedintheresidualblockofMobileNetv2toretainthecoordinateinformationinthechannelattention;Secondly,theimprovedMobileNetv2isusedtolightentheYOLOv4backbonenetwork,andthedepthwiseseparableconvolutionblockisusedinPANettoreducetheamountofcomputation;Then,ASFFadaptivefeaturefusionisusedtoimprovethePANetstructuretobalancetheinconsistencyofdifferentfeaturelayers.Finally,attentionisaddedtothefeaturefusionmoduletoincreasetheweightofthetargetinformation;andtheK-means++algorithmgeneratesnewaprioriboxclustercenters.Experimentsshowthattheweightfileisreducedby60%from136Mto54.5M,thenetworkvolumeisreducedby80%,andtheaccuracyreaches96.84%,loseonly0.46%accuracy...
