Seite - 183 - in Document Image Processing

J. Imaging 2017,3, 62 images;and(3) font,backgroundandlayoutarefinallyusedwiththetranscribedtexts toautomatically generate syntheticversionsof theoriginal ones. Thewholedatabase is composedof 93document imagescontaining18.240wordsand115.622characters. (a) (DB1)Contemporaryfrenchtypewrittendocument. (b) (DB2)Oldfrenchtypewrittendocument. (c) (DB3)OldfrenchManuscriptdocument. Figure11. Imagesextractedfromthedatabaseusedfor testingourpredictionalgorithm. (Left)original images. (Right) syntheticgenerated images. Toevaluate theOCRtext recognitionrate,weuse theLevenshteindistance (ametricmeasuring the difference between two strings) between the whole original transcribed text and the whole recognized text. Wecompute themeanof theLevenshteindistances for theNdocuments of each database. Using this Levenshtein distance, the difference between theOCR text recognition rate computedonreal imagesandtheonecomputedon“Loremipsum”version(Table2Column1and Column2) is, on average, only overestimated by 0.04. Thedifference between the realOCR rate and theone computedon the synthetic versions (Table 2Column1andColumn3) is, onaverage, onlyoverestimatedby0.03.Mostof thesuccessofdifferentexistingOCRpredictionmethods ([63–66]) arerelatedto thequalityandquantityof theneededgroundtruth.Ourpredictionmethodpresented hereprovidescomparableresultswith theones formthestateof theart. Table2.ComparisonbetweenOCRrecognitionratesobtainedonthreedifferentbooksoriginal images and their syntheticversions. Column1: OCRrecognition rateonoriginal images,Column2: OCR recognitionrateonsynthetic imagesgeneratedwithboththetextandthefont fromtheoriginal images, Column3:OCRrecognitionrateonsynthetic imagesgeneratedwith loremipsumrandomtextandthe font fromtheoriginal images. Original Image FontFrom SameText LoremText DB1 0.95 DB1 0.94 0.88 DB2 0.80 DB2 0.85 0.84 DB3 0.24 DB3 0.21 0.23 5.Conclusions DocCreator gives toDIAR researchers a simple and rapidway to extend existing document imagedatabasesor tocreatenewonesavoidingthe tedious taskofmanualgroundtruthgeneration. DocCreator embeds many fonts, backgrounds, meshes and realistic degradation models which, whencombined, result inan interestingcombinationofground-trutheddatabases. Theexperiments 183