Page - 5 - in Document Image Processing

J. Imaging 2018,4, 68 dependsontheaccurateregistrationof rectoandversosidesof thedocument,which isanon-trivial pre-processingstep. For a plausible restoration of documents with bleed-through, in addition to bleed-through identification,findingasuitablereplacementfortheaffectedpixels isalsoessential. Therestoredimage generated inmostof theabovemethods is eitherbinary,pseudo-binary (uniformbackgroundand varyingforegroundintensities),or textured(thebleed-throughregionsarereplacedwithanestimate of the localmeanbackground intensityorwith a randompattern). Anestimateof the localmean backgroundisusedin[6,18],butsuchmethodsaregoodformanuscriptswithareasonablysmooth backgroundwhileproducingvisibleartifacts fordocumentswithahighly texturedbackground. In [7], a random-fill inpaintingmethodissuggested toreplace thebleed-throughpixelswithbackground pixels randomlyselected fromtheneighbourhood. However, the randompixel selectionproduces salt andpepper like artifacts in regionswith largebleed-through. In [6,16], as apreliminary step, a “clean” background for the entire image is estimated, but this is usually a very laborious task. Inbleed-throughremoval, thedesiredrestoredimageis theonewheretheforegroundandbackground texture ispreservedasmuchaspossible. Instead,mostof thebleed-throughremovalmethodsusually concentrateonforegroundtextpreservation,neglectingthebackgroundtexture. Inorder toenhance the quality of the restored image, the identification of bleed-throughpixels and the estimationof a tenablereplacement for themshouldbeaddressedwithequalattention. Image inpainting,whichrefers tofilling inmissingorcorruptedregions inan image, isawell studiedandchallengingtopic incomputervisionandimageprocessing[19,20]. In image inpainting, the goal is to find an estimate for those regions in order to reconstruct a visually pleasant and consistent image[21]. Recently, sparserepresentationbasedimage inpaintingmethodsarereported withexquisite results [22,23]. Thesemethodsfindasparse linearcombinationforeach imagepatch usinganovercompletedictionary,andthenestimatethevalueofmissingpixels inthepatch. This linear sparserepresentation iscomputedadaptively,byusingaearneddictionaryandsparsecoefficients, trained on the image at hand.Adictionary learning basedmethod has been used for document image resolution enhancement [24], denoising [25], and restoration [26]. In addition to sparsity, non-localself-similarityisanothersignificantpropertyofnatural images[27,28].Anumberofnon-local regularization terms,exploiting thenon-local self-similarity,areemployedinsolving inverseproblems [29,30]. Fusingimagesparsitywithnon-localself-similarityproducesbetterresults inrecentlyreported image restoration techniques [31,32]. Theunderlying assumption in suchmethods is that similar patchesshare thesamedictionaryatoms. In thispaper,wepresenta two-stepmethodtorestoredocumentsaffectedbybleed-throughusing pre-registered recto andverso images. First, the bleed-throughpattern is selectively identified on bothsides; then, sparse image inpainting isused to findsuitable fill in for thebleed-throughpixels. Ingeneral,anyoff-the-shelfbleed-throughidentificationmethodscanbeusedinthefirst step.Here, weadoptthealgorithmdescribedin[33],whichissimpleandveryfast.Althoughefficientinlocatingthe bleed-throughpattern, themethodin[33] lacksaproperstrategytoreplacetheunwantedbleed-through pixels. Thesimplereplacementwith thepredominantbackgroundgray levelvaluecausesunpleasant imprintsof thebleed-throughpattern,visible in therestoredimage.Aninterpolationbasedinpainting techniqueforsuch imprints ispresented in [34],but the filled-inareasaremostlysmooth.Here,weuse asparse imagerepresentationbasedinpainting,withnon-local similarpatches, to findabefittingfill-in for thebleed-throughpixels. Thissparse inpaintingstep,whichconstitutes themaincontributionof thepaper,enhances thequalityof therestoredimageandpreserveswell thenaturalpaper textureand thetextstrokeappearance. Theoptimizationproblemofsparsepatch inpainting is formulatedusing thenon-local similarpatches, toaccount for theneighbourhoodconsistency,andorthogonalmatching pursuit (OMP) isusedtofindthesparseapproximation. Therestof thispaper isorganizedas follows. Thenextsectionbrieflyintroducessparse image representationanddictionary learning. Section3presents thenon-blindbleed-throughidentification method. Theproposed sparse image inpainting technique is described in Section 4. In Section 5, 5