Seite - 201 - in Document Image Processing

J. Imaging 2018,4, 32 Note that the sizeof the inputblock is set to 1×4 forProtocols 6.1 and3 (not 2×4), respectively. Tofine-tune theseparameterswe justpickoutasetof2000 labeled images fromAcTiV-R, inwhich 190areusedasavalidationset. Table6.Bestparameters for trainingthenetwork. Parameters Values MDLSTMSize 2,10and50 Feed-forwardSize 6and20 InputBlockSize 2×4 HiddenBlockSizes 1×4and1×4 Learnrate 10−4 Momentum 0.9 5.3. ExperimentalResults Several experiments have been conducted using the AcTiV-D and AcTiV-R subsets. These experimentscanbedivided into twocategories: Thefirstoneconcerns thecomparisonofoursystems with tworecentmethods. Thesecondcategoryaimsatanalyzingtheeffectof increasingthe training dataontheaccuracyof theLADItextdetector. 5.3.1.ComparisonwithOtherMethods As proof of concept of the proposed benchmark, we compare our systemswith two recent methods. ThefirstonewasproposedbyGaddouretal. [52] tobasicallydetectArabic texts innatural scene images. Themainsteps involvedare: • Pixel-colorclusteringusingk-means to formpairsof thresholds foreachRGBchannel. • Creationofbinarymapforeachpairof thresholds. • ExtractionofCCs. • Preliminaryfilteringaccordingto“areastability”criterion. • Secondfilteringbasedonasetof statisticalandgeometric rules. • Horizontalmergingof theremainingcomponents to formtextlines. Thesecondmethodwasput forwardbyIwataetal. [53] torecognizeartificialArabic text invideo frames. Itoperatesas follows: • Textlinesegmentation intowordsbythresholdinggapsbetweenCCs. • Over-segmentationofcharacters intoprimitivesegments. • Character recognition using 64-dimensional feature vector of chain code histogram and the modifiedquadraticdiscriminant function. • Word recognition by dynamic programming using total likelihood of characters as objective function. • Falseword reduction bymeasuring the average of the character likelihoods in aword and comparing it toapredefinedthreshold. Thedetectionsystemshavebeentrainedonthe training-set1ofTable4. Theevaluationhasbeen doneonthe test set for thedetectionandrecognitiontasks. Table7presentsevaluationresultsof the detectionprotocols in termsofprecision, recallandF-measure. Thebest resultsaremarkedinbold. TheLADI systemscoresbest for all protocolswith anF-measurebetween0.73 and0.85 forAllSD protocol (p4.4) andAljazeeraHDprotocol (p1) respectively. In contrast to theSysAthat represents a fully heuristic-basedmethod, the LADI system increased the F-measure by 11% for Protocol 1. ForProtocols 4.1, 4.2, 4.3 and4.4 (SDchannels), the results arehigher,withagainof, respectively, 11%,17%,14%and24%.Thisreflects theeffectivenessofusingamachine-learningsolutiontofilter the resultsgivenbytheSWTalgorithm.TheGaddosystemhasstrongfragmentationandmissdetection tendency asdepictedby its obtainednumerical results. Table 8presents evaluation results of the 201