Seite - 71 - in Document Image Processing

Journal of Imaging Article EfficientQuerySpecificDTWDistanceforDocument RetrievalwithUnlimitedVocabulary GattigorlaNagendar1,*,VireshRanjan2,GauravHarit 3 andC.V. Jawahar1 1 Center forVisual InformationTechnology, IIITHyderabad,Hyderabad500032, India; jawahar@iiit.ac.in 2 CSEDepartment,StonyBrookUniversity,StonyBrook,NY11794,USA;viresh.ranjan@stonybrook.edu 3 DepartmentofComputerScienceandEngineering, IIT Jodhpur, Jodhpur342037, India;gharit@iitj.ac.in * Correspondence: nagendar.g@research.iiit.ac.in Received: 31October2017;Accepted: 2February2018;Published: 8February2018 Abstract: In thispaper,weimprovetheperformanceof therecentlyproposedDirectQueryClassifier (DQC). The (DQC) isaclassifierbasedretrievalmethodandingeneral, suchmethodshavebeenshown tobesuperior to theOCR-basedsolutions forperformingretrieval inmanypracticaldocument image datasets. In (DQC), theclassifiersare trainedforasetof frequentqueriesandseamlesslyextendedfor therareandarbitraryqueries. Thisextends theclassifierbasedretrievalparadigmtoanunlimited numberofclasses (words)present ina language. The (DQC) requires indexingcut-portions (n-grams) of thewordimageandDTWdistancehasbeenusedfor indexing.However,DTWiscomputationally slowandtherefore limits theperformanceof the (DQC).WeintroducequeryspecificDTWdistance, which enables effective computation of global principal alignments for novel queries. Since the proposedqueryspecificDTWdistance isa linearapproximationof theDTWdistance, it enhances the performanceofthe (DQC).Unlikepreviousapproaches, theproposedqueryspecificDTWdistanceuses boththeclassmeanvectorsandthequery informationforcomputingtheglobalprincipalalignments for thequery. Since theproposedmethodcomputes theglobalprincipalalignmentsusingn-grams, itworkswell forbothfrequentandrarequeries.Wealsousequeryexpansion (QE) to further improve theperformanceof ourquery specific DTW. This also allowsus to seamlessly adapt our solution tonew fonts, styles andcollections. Wehavedemonstrated theutility of theproposed technique over3differentdatasets. TheproposedqueryspecificDTWperformswellcomparedtotheprevious DTWapproximations. Keywords:DTWdistance;queryclassifiers;wordspotting; indexing; retrieval 1. Introduction Retrievingrelevantdocuments(pages,paragraphsorwords) isacriticalcomponent ininformation retrieval solutionsassociatedwithdigital libraries. Theproblemhasbeen lookedat in twosettings: recognitionbased [1,2] likeOCRandrecognition free [3,4]. Mostof thepresentdaydigital libraries useOpticalCharacterRecognizers (OCR) for the recognitionofdigitizeddocuments and thereafter employa textbasedsolution for the information retrieval. ThoughOCRshavebecome thede facto preprocessingfor theretrieval, theyarerealizedas insufficient fordegradedbooks [5], incompatible for olderprintstyles[6],unavailableforspecializedscripts[7]andveryhardforhandwrittendocuments[8]. Evenforprintedbooks, commercialOCRsmayprovidehighlyunacceptableresults inpractice. Thebest commercialOCRscanonlygivewordaccuracyof90%onprintedbooks[4] inmoderndigital libraries. Thismeans that every 10thword in a book is not searchable. Recall of retrieval systems built on sucherroneoustext is thus limited.Recognitionfreeapproacheshavegainedinterest inrecentyears. Wordspotting[3] isapromisingmethodforrecognitionfreeretrieval. In thismethod,wordimagesare representedusingdifferent features (e.g.,Profiles, SIFT-BOW),andthefeaturesarecomparedwiththe helpofappropriatedistancemeasures (Euclidean,EarthMovers [9],DTW[10]).Wordspottinghas the J. Imaging 2018,4, 37 71 www.mdpi.com/journal/jimaging