Seite - 49 - in Document Image Processing

J. Imaging 2018,4, 57 ofLBP,all thenon-uniformpatternsaremarkedwiththesamelabel,whereas, foruniformpatterns, different labelsareused,one foreachpattern. This isperformedbecause ithasbeenobservedthat certainpatternsconstituteamajorportionofall texturefeatures.ULBPuses,M×(M−1)+3symbols to label thepatterns. 2.4. Rotation InvariantandUniformLBP(RIULBP) InRIULBP[22], thepatternsarechosensuchthat theyarebothrotation invariantanduniform. Similar toULBP,herealsoallnon-uniformrotation invariantpatternsareplaced inoneseparatebin. ThisvariantofLBPcanbeformulatedas RIULBP(M,R)(xcen,ycen)= { ∑Mn=1 f(In− Icen), ifU(RILBP(M,R)(xcen,ycen))≥2, M+1, otherwise. (7) Here, U(RILBP(M,R)(xcen,ycen))=( M ∑ n=2 |f(In− Icen)− f(In−1− Icen)|) + |f(IM− Icen)− f(I1− Icen)|. (8) 2.5. RobustandUniformLBP(RULBP) In the present work, we have proposed a minor but significant modification to Robust LBP (RLBP) [24] to develop RULBP. In RLBP, the argument of the function f(x) i.e., (In− Icen) (seeEquation(2)) is replacedwith (In− Icen− th),where thactsasa thresholdvalue. Thisessentially means that thevalueof Inhas tobegreater than thecenterpixel’sgrayvalue Icenbyanamount th toproducea1(seeFigure4). Thisdescriptor isdevisedwith the ideaof increasingtherobustness to negligiblechanges ingrayvalue. Therefore, theRLBPcanbeformallydefinedas follows: RLBP(M,R)(xcen,ycen)= M ∑ n=1 f(In− Icen− th)×2n−1. (9) In thiswork,wehavegivenanotionof setting thevalueof th for text/non-text separation in handwrittendocumentsandalsoincorporatedtheideaof ’uniformpattern’ inRLBPtodevelopRULBP. Figure 4. IllustrationofRLBPvaluegeneration for a 3× 3gray imagewindow,whereM=8and Radius=1.Here, thevalueof th=90. 2.5.1. Ideaof ‘UniformPattern’ To prove the effectiveness of LBP for texture classification [22], it has been shown that over 90 percent of the LBPs (generated using a segment of the image) present in a textured surface are ‘uniformpatterns’. Besides that, as ‘uniformpatterns’ consider a very limitednumber of 0/1 transition, theycanefficientlydetect the commonmicrofeatures like corner, edgeandspots. Thus, 49