Seite - 20 - in Document Image Processing

J. Imaging 2018,4, 27 The background-background probability is a function that needs to be optimized in the decision-making block, mapping background pixels (paper) from the original image ontowhite pixelsof thebinary image. Itdependsofall theparametersof theoriginal imagetexture, strengthof thebacktofront interference (simulatedbythecoefficientα),paper translucidity,etc. foreachRGB channel. Thus,onecanrepresent thisdependenceas: P(b/b)=f(α,R,G,B). (1) Theoptimal threshold tc* foreachchannel is calculated in thedecision-makingblock, the indexc canbeR,GorB,maximizingP(b/b): tc*=MaxP(b/b), (2) subject toagivencriterionP(f/f)≥M.ThecriterionusedherewasM=97%, that is atmost3%of the foregroundpixelsmaybe incorrectlymapped. During the trainingphase, the best tc*will be chosenfromthethreechannels,whichbestmaximizes theP(b/b) foreachof the images in the training set. Thematrixof co-occurrenceprobability is calculatedand thedecisionmaker chooses thebest binary image. Thedecision-makingblockwas trainedwith32,000synthetic images insuchawayto, givenareal imagetobebinarized, itfinds theoptimal thresholdparameters. 2.2.GeneratingSynthetic Images TheDecision-MakingBlockneedstrainingto“learn”about theoptimal thresholdparametersand thevalueof thekernel tobeusedinthebilateralfilter. Suchtrainingmustbedoneusingcontrolled imageswhicharesynthesizedtomimic thedifferentdegreesofback-to-front interference,paperaging, paper translucidity, etc. Figure4presents theblockdiagramfor thegenerationofsynthetic images. Twobinary imagesofdocumentsofdifferentnature (typed,handwrittenwithdifferentpens,printed, etc.) are taken: F—frontandV—verso (back). The front image isblurredwithaweakGaussianfilter to simulate thedigitalizationnoise [1], thehues thatappear inafterdocumentscanning. Figure4.BlockdiagramoftheschemeforthegenerationofsyntheticimagesfortheDecision-MakingBlock. 20