Seite - 66 - in Document Image Processing

J. Imaging 2018,4, 6 Figure5.Anexampleofarescoring lattice. 6. ExperimentResults Totrain theproposedholisticArabicOCRsystem,weuseda lexiconofaround356,000words selected from the newsdomainwith high coverage for theArabic Language. Using this lexicon, wegeneratedadatabaseof images for three fonts: SimplifiedArabic,TraditionalArabicandArabic Transparent, in300dpiwith fourdifferentsizes. To test the system, we used three different test datasets that represent different degrees ofchallenges: 1. Laserscannedtextdataset: Thisdataset iscomposedof1152singlewordstakenfromnewspaper articles andprinted in three fonts and fourdifferent sizes in two typesofqualities: cleanand firstcopy. 2. Recent computerizedbooksdataset: Adataset composedof10scannedpages fromdifferent recentcomputerizedbooks thatcontain2730words. 3. Oldun-computerizedbooks: Thisdataset consistsof10scannedpagescontain2276words from oldbooks thatare typewrittenwithnotwellknownfonts. Figure6 illustratessomeexamplesof thescannedimages. In thefirstexperiment,weevaluated our systemusing the laser scanned data set. Initially, we evaluated the systemon a single font. Thesystemwas trainedonasingle fontwithsinglesizebutwas testedonthesamefontwithdifferent sizes. Wedidn’t use the languagemodelwith this dataset as it consists of singlewords. Table 2 illustrates theWordRecognitionRate (WRR)results for thisexperiment. Figure6.Somesamplesof thescannedimages. 66