Lampung - a New Handwritten Character Benchmark: Database, Labeling and Recognition

12 Pages·2011·0.62 MB·English

Lampung - a New Handwritten Character Benchmark:
Database, Labeling and Recognition
Akmal Junaidi, Szil´ard Vajda, Gernot A. Fink
Computer Science Department, TU Dortmund, Germany
{akmal.junaidi,szilard.vajda,gernot.fink}@udo.edu
September 17, 2011
Overview of the talk: (cid:73) Labeling
(cid:73) Features
(cid:73) Introduction
(cid:73) Motivation (cid:73) Experiments
(cid:73) Script (cid:73) Conclusion
Motivation
New script:
(cid:73) lack of publications
(cid:73) no representative dataset
Cultural heritage
(cid:73) originated from Brahmi
script
(cid:73) preserving important
heritage
(cid:73) proof of script existence
AkmalJunaidi,Szil´ardVajda,GernotA.Fink
MultilingualOCR2011,Beijing,China Introduction Labeling Features Experiments Conclusion1
Lampung alphabet
Diacritics:
Characteristics:
Punctuation marks
Handwriting sample
(cid:73) not cursive
(cid:73) curve(s)
(cid:73) 20 letters
(cid:73) the name: Kaganga
AkmalJunaidi,Szil´ardVajda,GernotA.Fink
MultilingualOCR2011,Beijing,China Introduction Labeling Features Experiments Conclusion2
Semi-Automatic Labeling: An overview 1
1
Vajdaet.al,Semi-SupervisedEnsembleLearningApproachforCharacterLabelingwithMinimalHumanEffort,
ICDAR,2011
AkmalJunaidi,Szil´ardVajda,GernotA.Fink
MultilingualOCR2011,Beijing,China Introduction Labeling Features Experiments Conclusion3
Features
Water reservoir:
Structural and statistical:
(cid:73) top and bottom
(cid:73) branch points
(cid:73) gravity center
(cid:73) end points
(cid:73) size (volume)
(cid:73) pixel density
(cid:73) height and width
AkmalJunaidi,Szil´ardVajda,GernotA.Fink
MultilingualOCR2011,Beijing,China Introduction Labeling Features Experiments Conclusion4
Experiments
Dataset:
Classification: Neural network
(cid:73) fairy tales transcription
(cid:73) 82 docs. written by students
(cid:73) 35,193 character images
(cid:73) clustered to 11 classes
Composition:
(cid:73) 21,122 for training set (60%)
(cid:73) 10,547 for test set (30%)
(cid:73) 3,524 for validation set (10%)
Recognition result
Features #Training #Test Rec (%)
Branch points, end points, pixel density (BED) 21,122 10,547 93.2±0.48
Water reservoirs (WR) 21,122 10,547 91.3±0.54
BED and WR 21,122 10,547 94.3±0.44
AkmalJunaidi,Szil´ardVajda,GernotA.Fink
MultilingualOCR2011,Beijing,China Introduction Labeling Features Experiments Conclusion5
Misclassification
Variability in writing style
Different location of water reservoir
Unfiltered punctuation marks
Artifacts:
(cid:73) touching characters
(cid:73) character connected to
diacritic(s)
(cid:73) character connected to
punctuation mark(s)
AkmalJunaidi,Szil´ardVajda,GernotA.Fink
MultilingualOCR2011,Beijing,China Introduction Labeling Features Experiments Conclusion6
Conclusion
(cid:73) The Lampung:
(cid:73) scientific research challenge for handwritten recognition
(cid:73) preserving efforts of the Lampung as a cultural heritage
(cid:73) Semi-automatic labeling strategy: new approach
(cid:73) efficient labeling task for large dataset, minimize human
involvement
(cid:73) only 20% samples need to be relabeled
(cid:73) Water reservoir can effectively distinguish the Lampung
characters:
(cid:73) 91.3% recognition only based on water reservoir features
(cid:73) 94.3% recognition combining with branch points, end points, pixel
density
(cid:73) Lampung character dataset:
(cid:73) publicly available soon
(cid:73) preferably on TC11 website
AkmalJunaidi,Szil´ardVajda,GernotA.Fink
MultilingualOCR2011,Beijing,China Introduction Labeling Features Experiments Conclusion7
References I
[1] U.BhattacharyaandB.B.Chaudhuri.
DatabasesforResearchonRecognitionofHandwrittenCharactersofindianScripts.
InInternationalConferenceonDocumentAnalysisandRecognition,volume2,pages789–793,2005.
[2] B.B.ChaudhuriandS.Ghosh.
OrientationDetectionofMajorIndianScripts.
InProceedingsoftheInternationalWorkshoponMultilingualOCR,MOCR’09,pages8:1–8:7,NewYork,NY,
USA,2009.ACM.
[3] P.T.Daniels.
TheWorld’sWritingSystems.
OxfordUniversityPress,1996.
[4] D.Ghosh,T.Dube,andA.Shivaprasad.
ScriptRecognition:AReview.
IEEETrans.PatternAnal.Mach.Intell.,32:2142–2161,December2010.
[5] G.E.HintonandR.R.Salakhutdinov.
ReducingtheDimensionalityofDatawithNeuralNetworks.
Science,313(5786):504–507,July2006.
[6] M.S.Khorsheed.
RecognisingHandwrittenArabicManuscriptsUsingaSingleHiddenMarkovModel.
PatternRecogn.Lett.,24:2235–2242,October2003.
[7] L.I.Kuncheva.
CombiningPatternClassifiers:MethodsandAlgorithms.
Wiley-Interscience,2004.
AkmalJunaidi,Szil´ardVajda,GernotA.Fink
MultilingualOCR2011,Beijing,China Introduction Labeling Features Experiments Conclusion8
References II
[8] Y.LeCun,L.Bottou,Y.Bengio,andP.Haffner.
Gradient-BasedLearningAppliedtoDocumentRecognition.
InIntelligentSignalProcessing,pages306–351.IEEEPress,2001.
[9] C.-L.LiuandC.Y.Suen.
ANewBenchmarkontheRecognitionofHandwrittenBanglaandFarsiNumeralCharacters.
PatternRecognition,42:3287–3295,December2009.
[10] L.M.LorigoandV.Govindaraju.
OfflineArabicHandwritingRecognition:ASurvey.
IEEETrans.PatternAnal.Mach.Intell.,28:712–724,May2006.
[11] T.Mondal,U.Bhattacharya,S.K.Parui,K.Das,andV.Roy.
DatabaseGenerationandRecognitionofOnlineHandwrittenBanglaCharacters.
InProceedingsoftheInternationalWorkshoponMultilingualOCR,MOCR’09,pages9:1–9:6,NewYork,NY,
USA,2009.ACM.
[12] S.Mozaffari,H.E.Abed,V.M¨argner,K.Faez,andA.Amirshahi.
IfN/Farsi-Database:aDatabaseofFarsiHandwrittenCityNames.
InInternationalConferenceonFrontiersinHandwritingRecognition,2008.
[13] S.Mozaffari,K.Faez,F.Faradji,M.Ziaratban,andS.M.Golzan.
AComprehensiveIsolatedFarsi/ArabicCharacterDatabaseforHandwrittenOCRResearch.
InTenthInternationalWorkshoponFrontiersinHandwritingRecognition,LaBaule(France),2006.
[14] W.Niblack.
AnIntroductiontoDigitalImageProcessing.
StrandbergPublishingCompany,Birkeroed,Denmark,1985.
AkmalJunaidi,Szil´ardVajda,GernotA.Fink
MultilingualOCR2011,Beijing,China Introduction Labeling Features Experiments Conclusion9