Page - 191 - in Document Image Processing

J. Imaging 2018,4, 32 segmentationtask, thebestF-score,90%,wasobtainedbyMishraetal. [30]. Thealgorithmismainly basedontwosteps: aGMMrefinementusingstrokeandcolor featuresandagraphcutprocedure. TheKAISTdataset [31]consistsof3000 images taken in indoorandoutdoorscenes (seeFigure2d forexamples). This isamultilingualdataset,which includesEnglishandKoreantexts.KAISTcanbe usedforbothdetectionandsegmentationtasks,as itprovidesbinarymasks foreachcharacter in the image. The text segmentationalgorithmofZhuandZhang[32]outperformsexistingmethodsonthis datasetwithanF-scoreof88%.Themethodisbasedonsuperpixel clustering. First, anadaptiveSLIC textsuperpixelgenerationprocedure isperformed.Next,aDBSCAN-basedsuperpixelclustering is usedto fusestrokesuperpixels. Finally,astrokesuperpixelverificationprocess isapplied. TheNEOCRdataset [33] contains 659natural scene imageswithmulti-oriented texts of high variability (see Figure 2c for examples). This database is intended for scene text recognition and providedmultilingualevaluationenvironments,as it includes texts ineightEuropeanlanguages. In2016,Veitetal. [34]proposedadataset forEnglishscene textdetectionandrecognitioncalled COCO-Text. Thedataset isbasedontheMicrosoftCOCOdataset,whichcontains imagesofcomplex everydayscenes. Thebestresultonthisdataset (67.16%)wasobtainedbythewinnerof theCOCO-Text ICDAR2017competition[35].Note that theparticipatingmethodsonthiscompetitionwereranked basedontheirAverageprecision(AP)withanIntersectionoverUnion(IoU)of0.5. Recently,ChngandChan[36] introducedanewdataset,namelyTotal-text, forcurvedscene text detectionandrecognitionproblems. It contains1555scene imagesand9330annotatedwordswith threedifferent textorientations. Figure 3. Some examples of text detection systems [18–20] showing the evolution of this area of researchover tenyears. As forArabic language,majorcontributionshavealreadybeenmadeintheconventionalfield ofprintedandhandwrittenOCRsystems[7,10].Muchprogressofsuchsystemshasbeentriggered thanks to theavailabilityofpublicdatasets. Examples include the IFN/ENIT[37]andKHATT[38] datasets for offline handwriting recognition andwriter identification; theAPTI database [39] for printedwordrecognition;andtheADABdataset [40] thatworksononlinehandwritingrecognition. However,handlingArabic textdetectionandrecognition formultimediadocuments is limitedto veryfewstudies [41–43]. Table 1 presents commonly used datasets for text processing in images and videos, and summarizes their features in terms of text categories, sources, tasks, script, information of 191