Seite - 53 - in Document Image Processing

J. Imaging 2018,4, 57 areperformedtochoose theoptimalvalueof th. In thiswork,Weka3 [28], adataminingsoftware (UniversityofWaikato,Hamilton,NewZealand),hasbeenusedforclassificationandvisualization purpose. Thevaluesof theclassifiers’parametersused in thecurrentexperimentaregiven inTable1. Table1.Detailvaluesof theparametersusedbytheclassifiersunderconsideration. Classifier ParameterswithValues NB •Batchsize: 100•Normaldistributionfornumericattributes MLP •LearningRate for thebackpropagationalgorithm: 0.3 •MomentumRate: 0.2 •Numberofepochs to train through: 500 •LearningRate: 0.3 SMO •ComplexityconstantC:1 •ToleranceParameter: 1.0×10−3 •Epsilonforround-offerror: 1.0×10−12 •Therandomnumberseed: 1 K-NN •K:1•Batchsize: 100 RF •Batchsize: 100 •Minimumnumberof instancesper leaf: 1 •Minimumnumericclassvarianceproportionof trainvariance forsplit: 1.0×10−3 •Themaximumdepthof the tree: unlimited 4.3. PerformanceMetrics Theperformancesof theLBPvariantsaremeasuredusingthe followingconventionalmetrics: Recall= TP TP+FN , (11) Precision= TP TP+FP , (12) FM= 2×Recall×Precision Recall+Precision , (13) Accuracy= TP+TN Total number of samples ×100%. (14) InEquations (11)–(14),TP,FP,TNandFN represent truepositive, falsepositive, truenegative andfalsenegative, respectively. It is tobenotedthatall theexperimentsaredoneusing3-foldcross validationandthefinal resultsarecomputedafter takingtheaverageperformanceof the three folds. 5. ExperimentalResults Detailed results for each LBP based feature descriptors except RULBPwith each of the five classifiers for thecurrentdatabasearegiveninTable2. FromTable2, it canbeobservedthat theRF classifieroutperformsothers. Thus, classificationresults forRULBPwithdifferent thresholdvaluesare computedusingRFclassifieronly.Wealsosee that theRULBPoperatorgives thebest accuracy in classification,amongall theLBPvariantsconsidered.Detailedresultsdepictingtheperformanceof RULBPfordifferent thresholdsaregiven inTable3.Apictorial comparisonamongtheperformances of different LBPoperators usingRF classifier is given in Figure 8. Figure 9 shows the imageof a documentcontainingtextwritten inBanglaandclassifiedusingRULBP,whichgives thebest result amongallLBPvariants. Inaddition to this, agraphical comparisonof theperformanceofvariousLBP 53