Seite - 60 - in Document Image Processing

Journal of Imaging Article AHolisticTechniqueforanArabicOCRSystem FarhanM.A.Nashwan1,MohsenA.A.Rashwan1,HassaninM.Al-Barhamtoshy2, SherifM.Abdou3,*andAbdullahM.Moussa1 1 DepartmentofElectronicsandElectricalCommunications,CairoUniversity,Giza12613,Egypt; far_nash@hotmail.com(F.M.A.N.);mrashwan@rdi-eg.com(M.A.A.R.); a.m.moussa@ieee.org (A.M.M.) 2 FacultyofComputingandInformationTechnology,KingAbdulazizUniversity, Jeddah21589,SaudiArabia; hassanin@kau.edu.sa 3 FacultyofComputers&Information,CairoUniversity,Giza12613,Egypt * Correspondence: s.abdou@fci-cu.edu.eg;Tel.: +20-10-2661-4479 Received: 30October2017;Accepted: 22December2017;Published: 27December2017 Abstract:Analyticalbasedapproaches inOpticalCharacterRecognition (OCR)systemscanendurea significantamountofsegmentationerrors, especiallywhendealingwithcursive languagessuchas theArabic languagewith frequent overlapping between characters. Holistic based approaches that consider whole words as single units were introduced as an effective approach to avoid such segmentation errors. Still the main challenge for these approaches is their computation complexity, especiallywhendealingwith largevocabularyapplications. In thispaper,we introduce a computationally efficient, holisticArabicOCRsystem. A lexicon reductionapproachbasedon clusteringsimilar shapedwords isusedtoreducerecognition time.Usingglobalword levelDiscrete CosineTransform(DCT)basedfeatures incombinationwith localblockbasedfeatures,ourproposed approachmanaged to generalize for new font sizes thatwere not included in the training data. Evaluationresults for theapproachusingdifferent test sets frommodernandhistoricalArabicbooks arepromisingcomparedwithstateofartArabicOCRsystems. Keywords:ArabicOCRsystems;holisticOCRapproach;holisticOCRfeatures; lexiconreduction 1. Introduction Cursivescriptsrecognitionhastraditionallybeenhandledbytwomajorparadigms: asegmentation- basedanalytical approachandaword-basedholistic approach. In theanalytical approach, the input wordis treatedasasequenceofunits (usuallycharacters). Eachunit is thenindividuallyrecognized[1–4]. Thisapproachhasseveraldisadvantages.Thesegmentationofcursivewordsisachallengingtaskand anyerrors in thatprocesswill increase theerrors in the followingrecognitionstep. Also,manyof the usedfonts forcursivescriptsextensivelyuse ligatureswheretwoormorelettersare joinedasasingle glyph,whichcomplicates thecharacter levelsegmentation. Figure1showssomechallengingsamplesof Arabicwords. Figure1.SomeexamplesofArabicwords thatcontain ligatureswithmanuallysegmentedcharacters. Cursivelywrittenwordcannotberecognizedwithoutbeingsegmentedandcannotbesegmented withoutbeingrecognized[5]. Thisphenomenon,knownasSayre’sparadox,pushes thecommunity to J. Imaging 2018,4, 6 60 www.mdpi.com/journal/jimaging