Seite - 188 - in Document Image Processing

Journal of Imaging Article OpenDatasetsandToolsforArabicTextDetection andRecognitioninNewsVideoFrames OussamaZayene1,2,*,SamehMasmoudiTouj1, JeanHennebert 3,Rolf Ingold2 andNajouaEssoukriBenAmara1 1 LATISLab,NationalEngineeringSchoolofSousse (Eniso),UniversityofSousse,Sousse4054,Tunisia; samehmasmouditouj@yahoo.fr (S.M.T.);najoua.benamara@eniso.rnu.tn (N.E.B.A.) 2 DIVAGroup,Departmentof Informatics,UniversityofFribourg(Unifr),Fribourg1700,Switzerland; rolf.ingold@unifr.ch 3 ICoSys Institute,HES-SO,UniversityofAppliedSciences,Fribourg1705,Switzerland; jean.hennebert@hefr.ch * Correspondence: oussama.zayene@unifr.ch Received: 26November2017;Accepted: 26 January2018;Published: 31 January2018 Abstract:Recognizingtexts invideo ismorecomplexthaninotherenvironmentssuchasscanned documents. Video texts appear in various colors, unknown fonts and sizes, often affected by compression artifacts and lowquality. In contrast to Latin texts, there are nopublicly available datasetswhich cover all aspects of theArabicVideoOCRdomain. This paper describes a new well-definedandannotatedArabic-Text-in-VideodatasetcalledAcTiV2.0. Thedataset isdedicated especially tobuildingandevaluatingArabicvideotextdetectionandrecognitionsystems.AcTiV2.0 contains 189video clips serving as a rawmaterial for creating 4063key frames for thedetection task and 10,415 cropped text images for the recognition task. AcTiV 2.0 is alsodistributedwith its annotationandevaluation tools thataremadeopen-source for standardizationandvalidation purposes. Thispaperalsoreportsontheevaluationofseveral systemstestedunder theproposed detectionandrecognitionprotocols. Keywords:videotextdetection;videotext recognition;AcTiVdataset;ArabicVideoOCR 1. Introduction Broadcastnewsandpublic-affairsprogramsareaprominentsourceof informationthatprovides dailyupdatesonnationalandworldnews.Nowadays,TVnewscastersarchivea tremendousnumber ofnewsvideoclips thanks to therapidprogress inmassstorage technology.As thearchivesizegrows rapidly, themanualannotationofallvideoclipsbecomes impractical. Sincethe80s,researchinOCRtechniqueshasbeenanattractivefieldinthedocumentanalysisand recognitioncommunity. Priorworkhasaddressedspecificresearchproblemsthathaveborderedon printedandhandwritten texts inscanneddocuments. Recently, embeddedtext invideoshasreceived increasingattentionas itoftengivescrucial informationabout themediacontent [1–3].Newsvideos generally contain two types of texts [2]: scene text and artificial text (Figure 1). The first type is naturally recordedaspartof sceneduringvideocapturing, suchas trafficandshopsigns. Thesecond typeof text isartificiallysuperimposedonthevideoduringtheeditingprocess.Comparedwithscene text, theartificialoneusuallyprovidesbriefanddirectdescriptionofvideocontent,whichis important forautomaticbroadcastannotation. Typically,artificial text innewsvideo indicatesspeaker’sname, location,event information, scoresofamatch,etc. Therefore, in thiscontext,weparticularly focuson thiscategoryof text. Recognizingtext invideos,oftencalledVideoOCR[4], isanessential task inmanyapplications suchasnews indexingandretrieval [5], videocategorization, largearchivemanagingandspeaker J. Imaging 2018,4, 32 188 www.mdpi.com/journal/jimaging