Seite - 38 - in Document Image Processing

J. Imaging 2018,4, 80 3.4. PrintDB ThePrintDBconsistsoffiveprinteddocument images thatareartificiallyslantedoverarange from−45◦ to +45◦, yielding a total of 455 slanted, printed document images. The exact slant is predetermined,makingevaluationof the techniqueeasier andmoreprecise. Thedocumentswere madefrompartsof .pdffilestoensurepreciseslantvalues. Thepageswereselectedtoincludedifferent typeof text types (including sparewriting and single/double columns). All the textwas slanted, keepingtheoriginaldimensions ((1/5) *A4_height * (1/2) *A4_width). Figure9showsanexampleof thePrintDBafter theapplicationof the technique. Figure 9. Application of our slant removal technique on sample of PrintDB (left); corrected by 35◦ (right). 3.5. Set-Upof theTextRatioRParameter Theamount of text in thewindow isvery important inorder todetect the slant angle. Little text includes very little information,while toomuch textwould increase the computational cost. In [17], therateof10%wasusedasapprovedamountof text in theselectedwindows justbytestand trial.Here,moredetailedexperimentsareperformed.Thetext ratioofeachwindowiscountedand comparedto theslantdetectionerror. Theexperimentwasperformedforeverywindowonseveral imagesof thevalidationset: 5document images fromtheTrigraphSlantDB,1 fromtheWashington DB, 1 from theBH2MDB, and 20 from thePrintDB. Since the handwritten imageswere of high resolution, theprocedurewasvery timeconsumingand just fewof themwereused. On theother hand, the imagesofprintedtextwereallused. InFigure10, thecurveofsumofslantsquareerrors (SSE)withreference to the text ratio ispresentedforprintedandhandwritten text. KDQGZULWWHQ 3ULQW'%66( 7H[W 5DWLR 5 Figure10.Sumofsquareerrorsaccordingto text ratioR for theprinteddb(dotted line)andthe three handwrittendatabases (broken line). 38