Seite - 199 - in Document Image Processing

J. Imaging 2018,4, 32 WRR= #words_correctly_recognized #words (5) LRR= #lines_correctly_recognized #lines (6) Figure 11 showsanexample explaining the impact onCRRandWRRmetrics resulting from substitutionanddeletionerrors. Figure11.ExampleofCRRandWRRcomputationbasedonoutputerrors. It isworthnotingthat theproposedprotocolshelpusunderstandinghowgeneric is thesystem, i.e., if a systemperformswell forProtocols7and9(independentlyof theTVchannel). For instance, in theAcTiVCompcontest,weobservedthat someparticipatingsystemsperformwell inHDresolution only, someothers arequite generic (i.e., good inbothSDandHDresolutions). Other systemsare incompatiblewithaspecificresolution.Variousexamplesofusingtheseevaluationprotocolswillbe presented in thenextsection. 5.ApplicationofAcTiVDatasets Theproposeddatasetshavebeenusedtobuildandevaluate twosystemsforArabicvideotext detectionandrecognition. The textdetector isbasedonahybridapproachcomposedofCC-based heuristicphaseandamachine learningverificationprocedure. Therecognizersystemconsistsofa Multi-DimensionalRNNs(MDRNNs)[49]coupledwithaConnectionistTemporalClassification(CTC) layer [50]. 5.1. LADIDetector The LADI text detection system is based in our previous work [14,46], with new added enhancementsconsideringthecolorconsistencyofnear text regions.Ourtextdetectorrepresentsa hybridapproachconsistingof twostages: aCC-basedheuristic algorithmandamachine learning classification. Themain ideaof thissystemis tocombinetwotechniques: anadaptedversionof the SWTalgorithmandaconvolutionalauto-encoder (CAE).AsshowninFigure12, thefirst stagestarts withapreprocessingsteptodecreasenoiseandfinedetail. It thencomputes theedgemapandX&Y gradients fromtheprocessedframeusingCannyandSobeloperators, respectively.After that, theSWT operator isperformedas follow. - Gradientdirectiondp iscalculated,ateachedgepixelp,which is roughlyperpendicular to the strokeorientation. - Asearchray r= p+n∗dp (n>0) starting fromanedgepixelpalongthegradientdirectiondp is shotuntilwefindanotheredgepixelq. If these twoedgepixelshavenearlyoppositegradient orientations, theray isconsideredvalid.Allpixels inside this rayare labeledbythe length |p−q|. Thenext step is togroupadjacentpixels in theresultingSWTimage intoCCs. This isdoneby applying aflood-fill algorithmbasedon consistency in strokewidth and color. TheCCs are then filteredusingasetof simpleheuristic rulesconcerningtheCCsize,position,aspect-ratioandcolor uniformity. TheremainingCCsare iterativelymergedintowordsandtextlinesbasedonaproposed 199