2023年第37卷第4期测试技术学报Vol.37No.42023(总第160期)JOURNALOFTESTANDMEASUREMENTTECHNOLOGY(SumNo.160)文章编号:1671-7449(2023)04-0342-06基于Transformer的多尺度物体检测①侯越千,张丽红(山西大学物理电子工程学院,山西太原030006)摘要:目前,Transformer基本模型对同一场景内不同尺寸物体的检测能力不足,其主要原因为各层等尺度的输入嵌入无法提取跨尺度特征,导致网络不具备在不同尺度的特征之间建立交互的能力。基于此,提出一种基于Transformer的多尺度物体检测网络,该网络采用跨尺度嵌入层初步对图像特征进行嵌入处理;利用多分支空洞卷积对输入进行下采样,通过调整并行分支的膨胀率使该结构具有多样的感受野;然后,由残差自注意力模块对输出嵌入结果进行处理,为特征图的局部和全局信息构建联系,使注意力计算融入有效的多尺度语义信息,最终实现多尺度物体检测。模型在COCO等数据集上进行训练,实验结果表明该方法与其他物体检测方法相比具有显著优势。关键词:物体检测;多尺度;Transformer;注意力机制;空洞卷积中图分类号:TP391.4文献标识码:Adoi:10.3969/j.issn.1671-7449.2023.04.011Multi-ScaleObjectDetectionBasedonTransformerHOUYueqian,ZHANGLihong(CollegeofPhysicalandElectronicEngineering,ShanxiUniversity,Taiyuan030006,China)Abstract:ThecurrentTransformerbasicmodelisinadequatefordetectingobjectsofdifferentsizeswithinthesamescene.Themainreasonforthisisthattheequal-scaleinputembeddingofeachlayercannotextractcross-scalefeatures,resultinginanetworkthatdoesnothavetheabilitytoestablishin-teractionsbetweenfeaturesofdifferentscales.Inthispaper,weproposeaTransformer-basedmulti-scaleobjectdetectionnetwork,whichusescross-scaleembeddinglayerstoinitiallyembedimagefea-tures,inwhichtheinputisdownsampledusingmulti-branchnullconvolution,andthestructureismadetohavediversesensoryfieldsbyadjustingtheexpansionrateofparallelbranches.Theoutputembed-dingresultsarethenprocessedbytheresidualself-attentionmoduletoconstructlinksforlocalandglob-alinformationofthefeaturemap,sothattheattentioncalculationincorporateseffectivemulti-scalese-manticinformationandfinallyachievesmulti-scaleobjectdetection.ThemodelsaretrainedondatasetssuchasCOCO,andtheexperimentalresultssho...
