Seite - 113 - in Document Image Processing

J. Imaging 2018,4, 43 Figure12.Balinesemanuscriptwithbinarizedgroundtruth image. Figure13.Sundanesemanuscriptwithbinarizedgroundtruth image. 4.1.2. EvaluationMethod Followingourpreviouswork[24]andtheevaluationmethodfromtheICFHRcompetition[25], threemetricsofbinarizationevaluationthatwereusedin theDIBCO2009contest [37]areused in the binarization taskevaluation for thiswork. Those threemetricsareF-Measure (FM)(Equation(3)),Peak SNR (PSNR) (Equation(5)), andNegativeRateMetric (NRM)(Equation(8)). F-Measure (FM):FMisdefinedfromRecallandPrecision. Recall= TP FN+TP ×100 (1) Precision= TP FP+TP ×100 (2) TP, definedas truepositive, occurswhen the imagepixel is labeledas foregroundand theground truth isalso.FP,definedas falsepositive,occurswhenthe imagepixel is labeledas foregroundbut the groundtruth is labeledasbackground.FN,definedas falsenegative,occurswhenthe imagepixel is labeledasbackgroundbut thegroundtruth is labeledas foreground(Equations (1)and(2)). FM= 2×Recall×Precision Recall+Precision (3) AhigherF-measure indicatesabettermatch. PeakSNR(PSNR):PSNRiscalculatedfromMeanSquareError (MSE) (Equation(4)). MSE= M ∑ x=1 N ∑ y=1 (I1(x,y)− I2(x,y))2 M∗N (4) PSNR=10× log10( C2 MSE ), (5) whereC isdefinedas1, thedifferencebetween foregroundandbackgroundcolors in thecaseof a binary image.AhigherPSNRindicatesabettermatch. 113