Seite - 39 - in Document Image Processing

J. Imaging 2018,4, 80 Themainvolumeofwindows, as it is shown inFigure11, includes textup to20%. Since the printed imageswithslantareanartificialdatabase, theresultderivedfromhandwritten imageswere taken intoaccountasacommonR, that is14%text in the image is thecaseofwindowswithminimum detectionerror. Themajorityofwindows include textup to20%(seeFigure11). Since the slanted printed imagesareanartificialdatabase, theresultderivedfromhandwritten imageswere taken into accountasacommonR=14%correspondingto theminimumdetectionerror (seeFigure10). [ KDQGZULWWHQ 3ULQW'% 7H[W 5DWLR 5 ZLQGRZV Figure11.Numberofwindowsaccordingto text ratio in thewindow.The largeamountclose tozero corresponds toalmostemptywindows(background). 3.6. Set-Upof theHeightHof theWindow Inorder toestimate thebestnumber for thewindowheight, thewholevalidationsetwasused inourprogramwiththeratioequal to0.14,asderivedbythepreviousexperiment, forvaluesof the height from1to10mb(main-bodies). Thesumofsquareerrors (SSE)wasconsideredasevaluation measure. The results are shown inFigure 12. TheSSE is higher than inFigure 10, as here, all the document imagesof thevalidationsetwereused,contrary toFigure10where just fewof themwere considereddueto thecomputationalcostof thecorrespondingexperiment. As it isobvious inFigure12, thevalueof twomainbodiesseemstogive thebest results (smaller SSE)as (the1: onemainbody)wasconsideredextremelysmall incaseof lowresolution. Figure12. Sumof squareerrors forvariouswidthsandheightsof thedetectedwindows. Thebest choice isfive. 3.7. Set-Upof theWidthWof theWindow Inordertoestimatethebestnumberforthewindowwidth, thewholevalidationsetwasalsoused inourprogramwiththeratioR=0.14andH=2mbs, forvaluesof thewidthfrom1to10main-bodies. Sumofsquareerrors (SSE)wasconsideredasevaluationmeasure. TheresultsareshowninFigure12. TheSSEismuchhigher than inFigure10,ashere,all thedocument imagesof thevalidationsetwere 39