Seite - 47 - in Document Image Processing

J. Imaging 2018,4, 57 LetTbethe featurevector representingthe local texture: T= func( Icen, I0, I1,..., IM−1), where Icen and Ip for p∈ {0, 1, ...,M−1} representgrayvaluesof thecenterpixelandtheneighboring pixels, respectively. Toachievegrayscale invariance, the textureoperator ismodifiedtoconsider the dif ference in intensitiesof thecenterpixelanditsneighbors: T= func( I0− Icen, I1− Icen, ..., IM−1− Icen). Furthermore, toachievearobustnessagainst thescalingofgrayscale,only thesignsofdifference in intensitiesareconsidered: T= func( f(I0− Icen), f(I1− Icen), ..., f(IM−1− Icen)). Here, f(x)= { 1, ifx≥0, 0, ifx<0. (2) Finally, theLBPoperator, for thecenterpixel pcenhaving intensityvalue IcenwithMneighbors (X1,X2,...,XM)of intensities (I1, I2,..., IM), respectively, canbedefinedbelow: LBP(M,R)(xcen,ycen)= M ∑ n=1 f(In− Icen)×2n−1. (3) LBPcreatesanM-bit string.Hence, forM=8, thevaluesofLBP(M,R)(xcen,ycen)canvaryfrom0 to255. Theprocess isdepicted inFigure1. Figure 1. Illustration of LBP value generation for a 3× 3 gray image window, where M = 8, andradius=1. Inorder toefficientlyextract texture featuresofvariouscomplexities, theoriginalLBPoperator hasbeenmodifiedtogenerateanumberofvariants. 2.1. ImprovedLBP(ILBP) Themaindifferencebetween ILBP [23] and simpleLBP is that, insteadof the intensityof the centerpixel, themeanintensityvalueofall thepixels, includingthecenterpixel, isusedtofindthe intensitydifferenceduringbinarypatterncomputation. Inaddition to that,whilecomputingILBP, the intensityof thecenterpixel isalsocomparedwithmeanintensity. ILBPis formallydefinedas follows: ILBP(M,R)(xcen,ycen)= M−1 ∑ n=0 f(In− Imean)×2n+ f(Icen− Imean)×2M, (4) 47