基于CycleGAN和改进YOLOv8n的X光违禁品检测

打开文本图片集

关键词：图像检测；X光违禁品；CycleGAN；YOLOv8n模型；混合卷积注意力机制；CARAFE模块；BiFPN中图分类号：TN919.8-34；TP391.4 文献标识码：A 文章编号：1004-373X（2026）06-0094-08

X-raycontraband detection based on CycleGAN and improved YOLOv8n

ZUO Jiawei，WEI Peixu，GEChao （SchoolofElectricalEngineering，NorthChinaUniversityofScienceandTechnology，TangshanO63210，China）

Abstract：Inallsion totheproblemsoflowacuracy，slow detectionspeedandtoofewsamples inpublicdataset，an XrayimagecontrabanddetectionalgorithmbasedonCycleGANandimprovedYOLOv8nisproposed.Inthismodel，theCycleGAN model isused to generate more contraband images with security ΔX -raystyle characteristics，and establishenough samplesof contrabandimage.AccordingtothecharacteristicsofdiferentsizesofobjectsinX-raycontrabandimages，thehybrid convolutionalatention（HCA）mechanismisdesigned toefectivelyenhancetheabilityofYOLOv8ntoextractkeyfeaturesof objects.Inalusiontotheproblemofinsuficientfeatureinformationof securityX-raycontrabandimages，CARAFEmoduleis usedtoreplace theoriginalYOLOv8nupsampling mode，BiFPNnetwork isusedtoimprovetheneck network，andtheeffective featurefusionmethodand upsampling methodareused tomadeuptheinsuficientimagefeature information.Aimingatthe problemof parameter redundancyof YOLOv8n detection head，parameter sharingandDBB moduleareusedtoreduce the numberofmodelparametersandensurethedetectionaccuracyofthemodel.Thedetectionalgorithmwastestedontheexpanded SIXarydataset.TheexperimentalresultsshowthatthemAPvalueofthisalgorithmcanreach93.1，which is2.9%higherthan thatoftheoriginalmodel，whilethenumberofparametersisdecreasedby8.5%，maintainingthedetectionspeedoforiginal model.Theperformanceofthisalgoritissignificantimprovedcomparedtotheoriginalalgorithm，verifyngtheefectivenesof itsimprovement.

Keywords：image detection；X-raycontraband;CycleGAN；YOLOv8n model;hybridconvolutionalatention mechanism; CARAFEmodule;BiFPN

0 引言

违禁品检测是维护社会安全和保障交通运输稳定运行的重要措施，但是国内的安检方式主要依赖于安检员通过X光图像的人工识别，存在误检、效率慢等诸多问题。（剩余10933字）