融合Transformer注意力的舰船要害部位检测.pdfVIP

2842025，61（11）ComputerEngineeringandApplications计算机工程与应用

融合Transformer注意力的舰船要害部位检测

高兵，祝宇鸿+

吉林大学通信工程学院，长春130000

+通信作者E-mail：zhuyh@126.com

摘要：针对舰船及要害部位检测任务的需求，设计了SEC-YOLOv5（SwinTransformerC3EECBAM-YOLOv5）舰

船及要害部位检测算法。SEC-YOLOv5算法利用SwinTransformer改进网络主干部分C3模块，增强模型对语义特

征的提取能力，提升要害部位小目标的识别精度；使用CBAM注意力机制加权有效特征信息，提升特征信息利用率；

设计C3EE（C3easyeffect）模块替换瓶颈层所有C3模块，扩大模型感受野，丰富模型的梯度信息流。基于RS-ship

数据集验证所提算法，以原YOLOv5算法为基准进行逐模块的消融对比实验，实验结果表明，SEC-YOLOv5算法在

增加少量参数量的前提下，将平均精度值mAP@0.5提高5个百分点。

关键词：SwinTransfomer；舰船检测；卷积神经网络；深度学习

文献标志码：A中图分类号：TP391；TJ765.3doi：10.3778/j.issn.1002-8331.2402-0099

ShipCriticalPartsDetectionwithTransformerAttention

GAOBing,ZHUYuhong+

CollegeofCommunicationEngineering,JilinUniversity,Changchun130000,China

Abstract：Accordingtotherequirementsofshipsandkeyparts,thispaperdesignsthedetectionalgorithmofSEC-YOLOv5

shipsandkeyparts.TheSEC-YOLOv5algorithmusesSwinTransformertoimprovetheC3moduleofthebackbonenet-

work,enhancingtheabilitytoextractsemanticfeaturesandimprovingtheidentificationaccuracyofsmalltargetsinkey

parts.ItalsousestheCBAMattentionmechanismtoweighteffectivefeatureinformationandimprovetheutilizationrate

offeatureinformation.ThepaperdesignstheC3EEmoduletoreplaceallC3modulesinthebottlenecklayer,expandthe

feelingfieldofthemodel,andenrichthegradientinformationflowofthemodel.BasedontheRS-shipdataset,thepro-

posedalgorithmisverified,andtheoriginalYOLOv5algorithmisusedasthebenchmarkforthemoduleablationcompar-

isonexperiment.TheexperimentalresultsshowthattheSEC-YOLOv5algorithmincreasestheaverageaccuracyvalue

mAP@0.5by5percentagepointsonthepremiseofincreasingasmallnumberofparameters.

Keywords：SwinTransformer;shipdetection;convolu