Page - 64 - in Document Image Processing

J. Imaging 2018,4, 6 Figure3.DCTbasedFeatureExtraction. 3.2.DiscreteCosineTransform4-Blocks (DCT_4B) In this featureset,firstlywefindtheCentreofGravity (COG)of imageandmake itas thestarting point; inorder tocalculate thecentreofgravity, thehorizontalandvertical centremustbedetermined bythe followingequations: Cx= M(1,0) M(0,0) (2) Cy= M(0,1) M(0,0) (3) whereCx is the horizontal centre andCy the vertical centre of gravity and M(p,q) the geometrical momentsof rank p+q: Mpq=∑ x ∑ y ( x width )p( y height )q f(x,y). (4) Thexandydeterminethe imagewordpixels. Thedivisionofxandybythewidthandtheheight of the image, respectively, causes thegeometricalmoments tobenormalizedandbe invariant to the sizeof theword[18]. Thismethoduses featuresofCOGandDCTat thesametime, thefirstoneasan auxiliary feature todivide the image into fourpartsandapplythesecondfeatureDCToneachpart asawhole. This featureset isextractedandimplementedas follows: 1. Calculate the COG of the word image and make it as a starting point as explained in Equations (1)–(4). 2. Use theverticalandhorizontalCOGtodivide thewordimage into fourregions. 3. Apply theDCTtoeachpartof thewordimage. 4. PerformzigzagoperationontheDCTcoefficientsofeach imagepart toget thefirstN/4values thatcontainmostwordinformationonthatwordpart. 5. RepeatSteps3and4sequentially forall thewordparts, andthencombine themtogether to form thefeaturevectorof thewordimage. 3.3.HybridDCTandDCT_4B(DCT+DCT_4B) This featurecombines the twofeaturesDCTandDCT_4B. 4. LexicalReductionandClustering To reduce the computation time for searching the whole lexicon in the recognition phase, thesimilarshapewordsareclusteredtogether. Thewordsearch isperformedin twosteps. In thefirst one, thewordclusteror thenearestn-clustersaredeterminedthenthebestmatchingwordinside that clusterareselectedas therecognitionoutput. Forwordsclustering,weusedtheLBGalgorithm[19] to cluster thewords ineachgroupdependingonclosenessof thewordshapes fromthepointofviewof theusedfeatures. For theclusteringprocess,weusedthesameDCTandDCT_4Bfeatures thatweuse for thewordrecognitionphase. Tomeasure theaccuracyof theclusteringstep,andalso lexical reduction,weusedaclustering accuracy measure which counts the number of times the test word exists within the selected cluster/clustersper the testedwords. Foravocabularysizeofaround356,000wordsofSimplified Arabicfont(14pt.),wetestedtheclusteringaccuracyusingatestsetof3465wordsandacodebooksize 64