Seite - 82 - in Document Image Processing

J. Imaging 2018,4, 37 5.6. Results onGeorgeWashingtonDataset TheGeorgeWashington(GW)dataset [31]contains4894wordimages from1471wordclasses. This isoneof thepopulardataset forwordimages.WeappliedourproposedmethodofDQCusingQS DTWforwordretrievalontheGWdataset. Table6providescomparativeresults forsevenmethods. Experimentsarerepeatedfor100randomqueriesandtheaverageover theseresultsarereported in the table.Wecanobserve that for theDQCtheproposedQSDTWgivesbetterperformance thanDTW. Wecanalsoobserve that for thenearestneighborclassifier,QSDTWdistance isperformingslightly superior to theDTWdistanceandFastapproximateDTWdistance. Thesuperiority isbecauseof the principalalignmentswhicharequeryspecific. Table6.RetrievalperformanceontheGeorgeWashington(GW)dataset. TheDQCmakesuseof top 800frequentclasses for indexingthecut-portions. Dataset mAPUsingNearestNeighbour mAPUsingDQC DTW FastApprxDTW [20] QSDTW Euclidean sDTW FastDTW [30] QSDTW GW 0.51 0.50 0.52 0.32 0.62 0.63 0.70 5.7. Setting theHyperparameters Theproposedmethodhasfewhyperparameters, like the lengthof thecutportionandthenumber ofcutspecificprincipalalignments. For tuningtheseparameters,werandomlychoose100queries for eachdatasetandvalidate theperformanceover thesequeries.Queries includedin thevalidationset arenotusedforreportingthefinal results. InTable7,wereporttheeffectofvaryingthecutportionlengthonretrievalperformance. ThemAP score is less forsmallercutportion length. In thiscase, the learnedalignmentsarenotcapturingthe desiredcorrelations. Thishappensbecause theoccurrenceofsmallercutportions isveryfrequent in thewordimages. For lengthmorethan30, themAPisagaindecreased. This isbecausetheoccurrences of largercutportionsarerare.Cutportion lengths in therangeof10 to20givebetter results. In this case, thecutportionsaregoodenoughtoyieldglobalprincipalalignments thatcandistinguishthe differentwordimages. Table 7. The table shows the change in retrieval performancewith the change in the lengthof cut portionoverall thedatasets (D1,D2,D3).Here l is the lengthof thecutportion. l D1 D2 D3 1 0.81 0.78 0.7 10 0.86 0.83 0.74 20 0.86 0.82 0.75 30 0.82 0.77 0.72 Weassessedtheeffectofvaryingthenumberofcut-specificprincipalalignmentsontheretrieval performanceonthe threedatasetsandtheresultsaregiven inTable8. It is seenthat theperformance degrades forall thedatasetswhenthenumberofalignments is chosenas30. Thiscanbeattributedto someredundantalignmentsgetting includedin thesetofprincipalalignments. Increasing thenumber ofalignments from10to20 improvesperformancefordatasetD1,buthasnoeffectontheperformance fordatasetsD2andD3. Therefore,wecanconcludethatrestrictingthenumberofprincipalalignments in therange10 to20wouldgivegoodresults. Inallourexperiments,weset thenumberofcut-specific principalalignmentsas10. 82