欢迎来到天天文库
浏览记录
ID:36458090
大小:2.35 MB
页数:65页
时间:2019-05-10
《基于连通域的文本定位方法研究》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、浙江工业大学硕士学位论文基于连通域的文本定位方法研究姓名:嵇新浩申请学位级别:硕士专业:电子与通信工程指导教师:石旭刚20070401浙江T业大学工程硕上学位论文本文还提出了基于最小生成树方法将字符连通域联合成文本区域的方法。假设在同一文本区域的字符连通域具有类似的特征(相近的颜色和大小等)以及相近的距离。根据两个连通域特征的相似性与位置关系构建字符连通域之间一条边的权值。遍历所有的字符连通域,则将字符连通域集合构建成了一张无向带权值的图。使用最小生成树和一个阈值可以得到图像中的各个文本区域。关键词文本定位,字符识别,连通域分析,场景文本,最小生成树,BP神经网络Ⅱ浙江T业大学T程硕士学位论文
2、CoNNECTEDCoMPoNENTBASEDAPPRoACHFoRTEXTLoCALIZATIoNINIMAGESABSTRACTLocatingtextisrefertodetectingandlocatingtheareaofcharactersinimageswhichhavecomplexbackground.EffectivelocatingtextincomplexbackgroundimagecanextendtheapplicationofOCRtechnologysuchascontentbasedimageandvideoretrieval,Cal"platelocati
3、onandrecognition,etc.Locatingtextincomplexbackgroundimageshasbecomeaveryhotresearchissueindocumentanalysisandrecognitionarea.Inthispaper,weconductanexhaustivesurveyoftextlocationmethods,categorizethem,anddiscusstheadvantageanddisadvantageofthem.Thenweproposetextlocationalgorithmbasedonconnectedcompo
4、nent(CC)andneuralnetwork.Thismethodcaneffectivelydetecttextregionsinimagesandisrobusttothevariationofcharacter’Ssize,color,andfont.CCbasedtextlocationmethodiscomposedbyfoursteps.First,theinputimageissegmentedbyimprovedNiblackmethod.ThenCCanalysisisutilizedtogetCCs.ThesetofcandidateofcharacterCCisobt
5、ained.m浙江工业大学T程顾十学位论文Third,weextractallkindsoffeaturesofcomponent.Atlast,acascadeofthresholdclassifiersisusedtoclassifyCCsintocharacterCCsornon-characterCCs.BPNeuralNetworkiSintroducedintotheclassificationofCCs.nefeaturesofCCareusedastheinputofBPNeuralNetwork.TrainingsamplesisgotbyhandandfeedintoNeu
6、ralNetworktotraintheparameters.ThetrainedNeuralNetworkcanclassifyCCswhichthecascadeofthresholdclassifierscannotclassifyandimprovetheprecisionoftextlocation.Inthispaper,wealsouseMinimumSpanningTreetocombineCCsintotextregions.WesupposethatthecharacterCCsinthesametextregionhavesanlesizeandcolor,andthat
7、meyareneareachother.AccordingtothedistanceandsimilaritybetweentwoCCs,theweightoftheedgeisdefine.BycalculatingallcouplesinthesetofcharacterCCs,agraphisgot.MinimumSpanningTreeisdividedintosubsetsbasedon
此文档下载收益归作者所有