Seite - 40 - in Document Image Processing

J. Imaging 2018,4, 80 used,contrary toFigure10where just fewof themwereconsidereddueto thecomputationalcostof thecorrespondingexperiment. As it isobvious inFigure12, thevalueof justfivemain-bodiesseemstogivesignificantlybetter results (smallerSSE)which is farsmaller thanseven, initiallyusedin[17]. Thevaluesof1and2that givevery lowerrorprovedvery lowinthecaseof lowresolution,whilemorecharacterscouldconfuse thesystem. 3.8. Set-Upof theNumberMofFragments toUse Inorder toestimate thebestnumber for thewindowwidth, thewholevalidationsetwasalso usedinourprogramwiththeratioR=0.14,H=2mbs,andW=5mbs, forvaluesof theamountof the fragmentsMof4and5. More fragments couldbeused,however it is aproblemin thecaseof small imagesor imagesof lowresolution. Thesumofsquareerrors (SSE)wasconsideredasevaluation measure for thescenarios: • Four fragments,meanof the fragments: SSEontheevaluationset563 • Five fragments,meanof the fragments: SSEontheevaluationset513 • Five fragments,medianof the fragments: SSEontheevaluationset509 Obviously, themedianwasfinallyselectedas thebestcase. 3.9. ExperimentalResults on theDatabases Having specified the parameters above, several experiments were conducted for the four databases,oneachtest set. Inorder todemonstrate theflexibilityof theproposedtechnique, twomore slantdetectiontechniqueswerealsousedfor theslantdetectionpart,usingtheparameters thatwere specifiedbytheuseofour technique [9]. The technique [23]detects theslantbyusing themainpartof the texthavingremovedthehorizontalparts, ascenders, anddescenders. The technique [24]estimates theslantbyusingthepeaksof theslantedwords. The experimental results for all databases are given in Table 1 through RMSE. For the TrigraphSlantDB: the RMSE is given only for the normal slants (between−45 and +45 degrees) but for thebothestimators (Axel&Rolland). In theTrigraphSlant, eachwriterwasasked to force differentslants in twoof the fourdocuments, thesearenotpresentedsince theresultswerestrange duetounnatural slant>45degrees. Table1.Experimental results Database ProposedSlantDetection[9] SlantDetection[23] SlantDetection[24] TrigraphSlant (Axelestimat.) 7.08 8.30 6.99 TrigraphSlant (Rollandest.) 7.43 7.97 7.44 GeorgeWashingtonDB 3.44 6.53 3.40 BH2MDB 4.68 6.04 5.03 PrintDB 2.99 4.32 2.97 Therewereseveral significantdifferences in theRMSEvaluesbetweenthevariousDBs. Several reasons for thisare: • In theTrigraphSlant, thewriting ismodernandnotasuniformas in thehistoricaldocuments. Whenexaminedbyhumanestimators,astandarddeviationof2.45wasobserved. • GeorgeWashingtonDBofhistoricaldocuments ismoreuniform. • BH2Mpresentsmoredensitywhichmadeourcharactermainbodysizealgorithmfailmore times. • PrintDBincludesprintedtext that isartificiallyslanted,andtherefore isuniform. • Wepresent in the followingexperiments inorder toevaluate the improvementbroughtbyour slantdetection and removal techniqueondocument analysis and recognition tasks. We thus conduct recognition experiments on printed documents with an OCR, and word spotting 40