Seite - 182 - in Document Image Processing

J. Imaging 2017,3, 62 createdusing the 127 original images and transformedusing our 3Ddistortionmodel. The tests presented in [39,59] confirm the conclusion of [60] about the impact of the degradation level on re-training,either fora taskofcharacter recognitionor layoutextraction. 4.2.NewResults onPerformancePredictionUsingDocCreator Here,weshowwhetherDocCreatorcanbeuseful forperformancepredictionofexistingmethods. 4.2.1. Increase thePredictionRateofPredictiveBinarizationAlgorithm In [61],wehavepresented an algorithm topredict error rates of 11 binarizationmethods on givendocument imagesso that thebestbinarizationmethodisautomaticallychosenforanyimage dependingon its quality. Thismethod requires ground-trutheddata as input of the training step. TheDIBCOdatabase [62]wasused.However, theDIBCOdatabasecontainsonly36 images. Weproposehere to extend theoriginalDIBCOdatabasebyusing the inkdegradationmodel. Since theDIBCOdatabasecontains36 images,weextenditwith thesamenumberofsemi-synthetic document images. Thisextendeddataset is thenusedto train thepredictionmodelof [61]. Our retraining tests show that the use of this extended dataset allows one to increase the performanceof thepredictionmodelof [61].Moreprecisely, theerror rateof thepredictionmodel decreases (until it levelsoff)whenthenumberofsemi-synthetic images in the trainingset increases. Onaverage, theerrorratedropsofabout15%comparedwithusingonlyreal images in the training set. The error rate convergeswhen theproportionof semi-synthetic images is around50%of the trainingset. 4.2.2. PredictOCRRecognitionRateUsingSynthetic Images Many on-line digital libraries propose a text search engine. To this end, the textwithin the document imageshas tobe transcribed.DependingontheOCRrecognitionratequality, threeoptions areavailable: (1)directlyuse theOCRresultwhentherecognitionrate isclose to100%; (2)manually correcttheOCRresultwhentheautomatictranscriptiongives“acceptablequality”,or(3)doacomplete manual transcription(oftenquiteexpansive).Asaconsequence, it isvery important tobeawareof theOCRrecognitionratebeforedecidingbetweenoneof these threesolutions. Theamountof recent publicationsonthis subject ([63–66]) reflects thescientific interest inpredictingOCRsrecognitionrate. Weproposehere tousesynthetic images topredict theOCRrateofadigitizedbookas follows: (1) font, backgroundand layoutareextracted fromoriginal images (withmethodsdescribed in II). It isnoteworthy tomention that the fontswereextracted thoroughly, inparticular to includeeven characters not recognized by theOCR, or even to adjustmargins of correctly labeled characters. (2)AnadaptedLoremipsumtext is randomlygeneratedandusedtocreatesynthetic imageswith the fontandbackgroundpreviouslyextracted. ThisadaptedLoremipsumisgeneratedwithaccentuated characters (é, à, ù, etc.) and old characters (ff, fi, s, fl,ffi) if the original text contains such characters. Generating such characters is important tohave a representativedataset for fairOCR testing. As a result, images like the one presented in Figure 2 are generatedwith the associated XMLgroundtruth. (3)AnOCR(Tesseract) isfinallyusedtorecognize the textonthesesynthesized images. This text iscomparedwith theLoremipsumgroundtruth text,givinganOCRrecognition rate.Weconsider that this recognitionrate isapredictionof theOCRrate if theOCRsoftwarewas appliedonoriginal images. Table2Column1provides theaverageOCRrecognitionrateobtained ontheoriginal images,Table2Column3refers to theaverageOCRratescomputedonthesynthetic “Loremipsum”imagesversions. Wealsopropose toevaluate thecapacityofourmethodtocorrectlypredict theOCRrecognition ratebycomparingoriginal imageswith their syntheticversiongeneratedwithexactly thesametext (seeFigure11tocomparetheoriginal imagesandtheirsyntheticversions). These imagesaregenerated following this protocol: (1) pages from three books (2 typewritten and 1manuscript book) have beenmanually transcribed; (2) font,backgroundandlayoutareautomaticallyextractedfromoriginal 182