Seite - 73 - in Document Image Processing

J. Imaging 2018,4, 37 therearesimilaritiesbetweenthe topalignments (least costalignments)ofdifferentpairsof sequences. In [20], theauthorsexploredthesesimilaritiesby learningasmall setofglobalprincipalalignments fromthegivendata,whichcapturesall thepossiblecorrelations in thedata. Theseglobalprincipal alignments are then used to compute the DTW distance for the new test sequences. Since these methods [19,20] avoid the computationof optimal alignments, these are computationally efficient comparedtonaiveDTWdistance. ThefastapproximateDTWdistancecanbeusedforefficient indexing inDQCclassifier.However, itgivessub-optimal results. Forbest results, itneedsqueryspecificglobal principalalignments. In thispaper,weintroducequeryspecificDTWdistance,whichenables thedirect designofglobalprincipalalignments fornovelqueries. Globalprincipalalignmentsarecomputed forasetof frequentclassesandseamlesslyextendedfor therareandarbitraryqueries,asandwhen required,withoutusing languagespecificknowledge. This isadistinctadvantageoveranOCRengine, which is difficult to adapt to varied fonts andnoisy images andwould require language specific knowledgetogeneratepossiblehypotheses foroutofvocabularywords.Moreover,anOCRenginecan respondtoawordimagequeryonlybyfirstconverting it into text,which isagainpronetorecognition errors. In [21,22], deep learning frameworksareused forword spotting. In [23], a attributebased learningmodelPHOC ispresentedforwordspotting. In trainingphase,eachwordimageis tobegiven with its transcription. Bothword imagefeaturevectorsandits transcriptionsareusedtocreate the PHOCrepresentation.AnSVMis learnedforeachattribute in this representation.Ourapproachbears similaritywiththePHOCrepresentationbasedwordspotting[23]. Inthissense,boththeapproachesare designedforhandlingout-of-vocabularyqueries.Ourworktakesadvantageofgranulardescription atngrams (cut-portion) level. This somewhat resembles thearrangementof charactersused in the PHOCencoding.However, trainingefforts for PHOCaresubstantialwitha largenumberofclassifiers (604classifiers)beingtrainedandrequirescompletedata for training,which ishugefor largedatasets. Inourwork, the amountof trainingdata is restricted toonly frequent classes,which ismuch less comparedtoPHOC. Further,PHOCrequireslabels intheformoftranscriptions,whereasinourworkthe labelsneednotbe transcriptions. Inaddition, PHOC is languagedependent [24]andit isverydifficult toapplyoverdifferent languages. Themethodproposed in thispaper is language independent; it can beappliedtoanylanguage. Thepaper isorganizedas follows. Thenextsectiondescribes theDirectqueryclassifier (DQC). Fast approximationof (DTW) distance is discussed in Section 3. Thequery specific DTWdistance ispresented inSection4. Experimental settingsandresults arediscussed inSection5, followedby concludingremarks inSection6. 2.DirectQueryClassifier (DQC) In[18],Ranjanetal. proposedDirectQueryClassifier (DQC),whichisaone-shot learningscheme for dynamically synthesizing classifiers for novel queries. Themain idea is to compute an SVM classifier for thequeryclassusing theclassifiersobtainedfromthe frequentclassesof thedatabase. Thenumberofpossiblewords ina languagecouldbevery largeanditwouldbepracticallydifficult to buildaclassifier foreachof thewords.However, all thesewordscomefromasmall setofn-grams. Thewordscorrespondingto the frequentqueriesareexpectedtocontain then-gramsthatcover the fullvocabulary. ExemplarSVMclassifiersarecomputedfor the frequentqueries (wordclasses)and thenappropriatelyconcatenatedtocreatenovelclassifiers for therarequeries.However, thisprocess has its challengesdueto (i) Variationsduetonatureofscriptandwritingstyle, (ii) Classifiers forsmallerngramscouldbenoisy. Theauthorsaddress these limitationsbybuildingtheSVMclassifiers formost frequentqueries anduseclassifier synthesisonly for rarequeries. This improves itsoverallperformance. Theyuse Query Expansion (QE) for further improving the performance. An overviewof the direct query classifier isgiven in the followingsections. 73