Seite - 63 - in Document Image Processing

J. Imaging 2018,4, 6 Figure2.BlockDiagramof theHolisticOCRSystem. 3. FeatureExtraction Themainconceptof theproposedalgorithmisbasedontheproperty that theDCTtransform compressed image is adecompositionvectorwhich canuniquely represent the input image tobe correctly reconstructed later at a decompression stage. In this work, the first 100–200 2D-DCT coefficients are used as word features that provide good approximation about the word image information. Inoursystem, three featureswereexperimented. Those featuresare: DiscreteCosine Transforms(DCT),DiscreteCosineTransforms4-Blocks(DCT_4B),andafeaturewhichisacombination ofDCTandDCT_4B. 3.1.DiscreteCosineTransform(DCT) TheDCTfeatures inoursystemareextractedvia twodimensionalDCT.Thetwodimensional DCTofanM×N imagef(x,y) isdefinedas follows: T(u,v)= 1√ MN CuCv M−1 ∑ x=0 N−1 ∑ y=0 f(x,y)cos( (2x+1)uπ 2M )cos( (2y+1)vπ 2N ) (1) where0≤ x≤M−1,0≤ y≤N−1 Cu= { 1√ M ,x=0 2√ M ,1≤ x≤M−1 , Cv= { 1√ N ,y=0 2√ N ,1≤ y≤N−1 . After applyingDCT to thewholeword image, the features are extracted in avector formby usingthemostsignificantDCTcoefficients. Thesteps involved inDCTfeatureextractionasshownin Figure3are: 1. Apply theDCTtothewholewordimage. 2. PerformzigzagoperationontheDCTcoefficients Idct. Thezigzagmatrix Iz isa rowvectormatrixcontaininghighfrequencycoefficients in itsfirstN values thatcontainmostwordinformation. This formsfeaturesvector fdct foreachword. 63