Page - 62 - in Document Image Processing

J. Imaging 2018,4, 6 trainedbytypewrittenArabicwords infivefontswithsize14pointsandlexiconsizeof252words. Vector quantizationwas used tomap each feature vector to the closest symbol in the codebook. Themultiple recognitionhypotheses (N-bestwordlattice)of that systemachieveda97.65%accuracy. Also, theholisticapproachwassuccessfullyusedonthesubwordlevel.NasrollahiandEbrahimi [15] presentedanapproachtoofflineOCRforprintedPersiansubwordsusingwaveletpacket transform. Theproposedtechniqueextractedfont invariantandsize invariant features fromdifferentsubwords of four fontsandthreesizesandcompressedthemusingPrincipalComponentAnalysis (PCA).When testedonasubsetof2000wordsofprintedPersian textdocuments, that systemachievedanaccuracy of97.9%. In a later work [16], Slimane et al. organized the ICDAR2013 competition on multi-font and multi-size digitally represented Arabic text. The main characteristic of the winner system, SiemenssystemsubmittedbyMarc-PeterSchambachet al.,was theusingofa threehidden layers neuralnetwork, that transformsatwo-dimensionalpixelplane intoasequenceofclassprobabilities. thesystemhavebeenappliedonasubsetof theAPTIdataset [17]andmanagedtoachieveanaccuracy over99%. While theholisticapproachavoids thechallengingsegmentationtaskofArabiccursivescripts, it still has another challengeofdealingwith large lexicon sizeofArabicwords. As thenumberof words inthe lexicongrows, therecognitiontaskbecomesmorecomputationallyexpensive.Mostof the previouslyproposedholisticbasedArabicOCRsystemstestedwithsmall sizevocabularies,but this is notpractical forArabicasamorphologicallyrich languagewithahugevocabularysize. In this paper,wepropose a computationally efficient holisticArabicOCRsystem for a large vocabularysize. For thesakeofapracticalapproach,a lexiconreductiontechniquebasedonclustering thesimilarshapewords isusedtominimize thewordrecognition time. Theproposedsystemutilizes ahybridofseveralholistic features thatcombineglobalwordlevelDCT-basedfeaturesandlocalblock basedfeatures.Usingthese typesof features, thesystemmanages toachieveOmni-fontperformance with fontandsize independence.Also, thepresentedsystemhasaflexiblearchitecture for integrating languagemodellingconstraintsbyusingasecondrescoringpass for the topn-bestwordhypotheses. This rescoringoperationprovidedasignificantenhancement in therecognitionaccuracyof thesystem. Therestof thepaper isorganizedas follows. Section2 includesadescriptionfor theproposedholistic OCR system. The holistic DCT features used are described in Section 3. The developed lexicon reductiontechnique is illustrated inSection4. Section5describes the languagerescoringprocessused bythesystem. Section6presentssystemevaluationresultsandperformancecomparisonwithstateof art commercialArabicOCRsystems. Thefinalconclusionsandprospects for futureworkare included inSection7. 2. SystemDescription ThedevelopedholisticOCRsystemconsistsof twomodules. Thefirstone is the trainingmodule where the holistic features are extracted from the training set of theword images. The extracted featuresareusedtobuild thesetof clustersof similarwordshapes. Thegeneratedwords’ clustersand theirextractedfeaturesrepresent theknowledgebase that isusedin therecognitionphase. Thesecond module is therecognitionmodule. In thatmodule,afterapplyingthepreprocessingoperationsonthe input image, thedetectedtextblocksaresegmentedintolinesandwords. Thefeaturesareextractedfor eachwordimagethenthewordclusterorbest-nclusters, thathave theminimumEuclideandistance with the test imagevector,areassigned. Thegeneratedwordlist fromtheselectedcluster isusedto constructawordlattice for thepossible recognitionhypothesesof thewhole line. Thiswordlattice is rescoredusingn-gramlanguagemodel toget thebest recognitionhypothesis. Figure2showsthe blockdiagramof theproposedholisticOCRsystem. 62