基于小波神经网络的交通标志识别研究

《基于小波神经网络的交通标志识别研究》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、Y976543”*{—；—一——————————一声％———————————辛束文通大学硕士学位论文基于小波神经网络的交通标志识别研究学位申请人：胡晓燕学科专业：交通信息工程及控制指导教师t蒋先刚教授答辩委员会主席碱涝"譬辩日期：2柙6-6·jTRAFFlCESlGNSRECOGNIT10NRESEARCHBASEDONWAVELETNEURALNETWORKABSTRACTAsweallknown,imagerecognitionisallimportantbranchofpattemrecognition．Throughfewdecades,ithasbeenapplieds

2、uccessfullyinthemilitary,spaceexploration，medicalscienceandpo瓯etc．Soithasgreatimportanceandpracticalvalue．ThepurposeofthisthesisiStoutilizethepredominanceofthewaveletneuralnetworkfortherecognitionofimages．T}lispaperrevolvesaroundthecentraltaskofimageidentificatiomItismainlyaboutcollectingandp

3、reprocessingtheofi百naldataoftargetimages，methodsofinvariablefeatureextractionandtheidentificationtechnologyofthewaveletneuralnetwork．Attheimageprepmcessingpart，wesmoothanddenoisethetargetimagefirstly,thenwedetectitsedgeandbuildupit,finally,inordertoeliminatetheeffectofthetranslation,sealing，s

4、kewingandrotationontherecognitionresult．Weproposedamethodtonormalizethetargetimage，whichinvolvedinthefeatureextractionpart．Duringthefeatureextractionpart，weiisemomentalgorithmanddiscussnormalmoment，waveletmomentindetail．Thetraditionalmomentinvariantshavethedefect：thesemomentsarethewholefeatur

5、escalculatedfromthewholeimagespace，whichareapttodisturbedbynoise．Aimattheabovedefeet,anewmomentinvanantwaveletmomentispresented,whichapplywaveletanalysistomomentmvariant．ThuswaveletmomentpossesstheimageobjectSinvarianttotranslation,sealingandrotation．Byusingwaveletmomentinvariant，notonlythelo

6、calfeatureofimageobjectisobtained,butalsothedeseriptionabilityforthefinefeatureofimageconstructisimpmved．Thusthehigherrecognitionrateisobtained,especialtosimilarimages．Intherecognitionpart,waveletmomentsareusedastheimagefeatures，thenthefeaturesabstractedareoptimizedandfinallythefeaturesoptimi

7、zeda托combinedwithBPneuralnetworkandwaveletneuralnetworkclassifiertomakeobjectimagerecognitiomInordertoellslll-everacityoftherecognitionresult,thispaperuseclassesoftrafficsignsfortrainingandtesting．1rI嵋experimentalresultsdemonstratethatcompare