基于CNN与Transformer相结合的工业零件缺陷检测

打开文本图片集

中图分类号：TN911.73-34 文献标识码：A 文章编号：1004-373X（2025）15-0119-08

Industrial parts defect detection based on combination of CNN and Transformer

XIAXinghua1²，LIJiaying1，²，HANZhonghua1，2 （1.Schoolof ComputerScience and Engineering，Shenyang Jianzhu University，Shenyang11O168，China; 2.SchoolofElectrical and Control Engineering，ShenyangJianzhu University，Shenyang1o168，China）

Abstract：Intheprocessof industrialproduction，therewillbecracks，stainsandcreasesinparts，soitisdificult tolearn boththeglobalinformatioandedgedetailsofimagesatthesametimebyasingleneuralnetwork，anditfilstodetectdefects accurately.Therefore，thispapercombinestheadvantagesofCNN'slocalfeatureextractionwiththeTransformer'sstrongabilityto capturetheglobaldependencies，extractsimagefeaturesbythetwobranchesofResNet34andSwinTransformer，respectively， introducesatention mechanism，enhancesattentioninspaceandchanneldimensions，andotainstheoutputoffeaturefusionby splicing，ostaltsifgobldloalifoaiallcoindotsulaleoiosi of UNetstructureandthefinalsemanticsegmenationresultsarebtainedbyup-sampling，splicingandpixel-by-pielclaificatio Thesegmentationresultmapsclearlyshowthedefectlocationsintheimages，thusthedefectdetectionofindustrialpartsisrealized. Bycomparativeexperiments，theIUresultsoftheproposedalgorithmontheMVTecADdatasetof industrialproductionpartsare muchbeterthanthoseoftheothersemanticsegmentationalgorithms，andthesegmentationresultmapsare ingodagreementwith thelabelresultsof the sample images，which fullverify the efectiveness and practicabilityof the proposed algorithm.

Keywords：deep learning; industrial parts; defect detection; CNN; Transformer; feature extraction

0 引言

在工业零件生产过程中，部分零件会出现裂痕、错位、污渍、折痕、异物与缺损等缺陷，为保证工业零件生产的合格率，进行缺陷检测十分重要。（剩余7246字）