融合空洞卷积的轻量化目标检测①李洋,苟刚(贵州大学计算机科学与技术学院公共大数据国家重点实验室,贵阳550025)通信作者:苟刚,E-mail:6706605@qq.com摘要:为了轻量化模型,便于移动端设备的嵌入,对YOLOv4网络进行了改进.首先,用MobileNetV3作为主干网络,并使用深度可分离卷积替换加强特征提取网络的普通卷积,降低模型参数量;其次,在104×104特征图输出时融合空洞率为2的空洞卷积,与52×52的特征层进行特征融合,获取更多的语义信息和位置信息,细化特征提取能力,提升模型对极小目标的检测性能;最后,将原来的池化层使用3个5×5的Maxpool进行串联,减少计算量,提升检测速度.实验结果表明,在华为云2020数据集上,改进算法的mAP比YM算法提高了2.33%,在公共数据集VOC07+12上,mAP提高了3.12%,FPS比原来的YOLOv4算法提高了一倍多,参数量降低至原来的18%,证明了改进算法的有效性.关键词:MobileNetV3;YOLOv4;空洞卷积;轻量化;深度可分离卷积引用格式:李洋,苟刚.融合空洞卷积的轻量化目标检测.计算机系统应用,2023,32(2):379–386.http://www.c-s-a.org.cn/1003-3254/8975.htmlLightweightTargetDetectionBasedonDilatedConvolutionLIYang,GOUGang(StateKeyLaboratoryofPublicBigData,CollegeofComputerScienceandTechnology,GuizhouUniversity,Guiyang550025,China)Abstract:Inordertomakethemodellightweightandfacilitatetheembeddingofmobiledevices,theYOLOv4networkisimproved.Firstly,MobileNetV3isusedasthebackbonenetwork,andadeepseparableconvolutionisadoptedtoreplacetheordinaryconvolutionofanenhancedfeatureextractionnetwork,soastoreducethenumberofmodelparameters.Secondly,whenthefeaturemapwithasizeof104×104isoutput,thedilatedconvolutionwithadilatedrateof2isfused,anditisthenfusedwithafeaturelayerwithasizeof52×52,soastoobtainmoresemanticandlocationinformation,whichcanrefinethefeatureextractionabilityandimprovethedetectionperformanceofthemodelforminimaltargets.Finally,theoriginalpoolinglayerisconnectedinserieswiththreeMaxpoolswithasizeof5×5toreducethecomputationalloadandimprovethedetectionspeed.TheexperimentalresultsshowthatonHuaweiCloud2020dataset,themAPoftheimprovedalgorithmisimprovedby2.33%comparedwiththeYMalgorithm,andonthepublicdatasetVOC07+12,themAPisimprovedby3.12%,andtheFP...
