Page - 52 - in Document Image Processing
Image of the Page - 52 -
Text of the Page - 52 -
J. Imaging 2018,4, 57
followed by details of the parameter values used by the classifiers. At the end, we present the
evaluationmetricsusedin theexperiment.
4.1.DatabasePreparation
It is foundthat theunavailabilityofastandarddatabasemaybeoneof thepossiblereasons for
slowprogress insomeresearchareas, suchas text/non-text separationfromhandwrittendocuments
in spite of their importance. Keeping this fact inmind, in thepresentwork, a database has been
developedthatconsistsof104handwrittenengineering labcopiesandclassnotescollectedfroman
engineering college. These copies include textual contents alongwithavaryingnumberof tables,
graphiccomponentsandsomeprintedtexts.Alloftheselabcopiesarewrittenbymorethan20students
fromdifferentengineeringandsciencestreams. Theageof thewritersvary from18to24. Pleasenote
thatall thesecopiesarewritteneither inEnglishorBangla. Thecollecteddocumentsarescannedin
300DPI(Dotsper inch)usingaflatbedscannerandthenthesescannedcopiesarestoredas24bit ’BMP’
files.Asample imagefromthecurrentdatabase is showninFigure7aandthecorrespondingground
truthimageisshowninFigure7b. Inthiswork, fromthose104handwrittenpages,atotalof66,058CCs
areextracted,outofwhich25,011are textcomponentsand41,047arenon-textcomponents.
(a) (b)
Figure7. (a) sample imagefromourdataset; (b)groundtruthof thegiven image(here, redrepresents
textandbluerepresentsnon-textcomponents).
4.2. Classifiers
Forclassificationof theextractedCCs,fivewell-knownclassifiersareused in thiswork,namely,
NaïveBayes (NB),Multi-layerperceptron (MLP),K-nearestneighbor (K-NN),Randomforest (RF)
andSupportVectorMachine (SVM). In thecurrentexperimental setup,performancesofSimpleLBP,
ILBP,RILBP,ULBPandRIULBPdescriptorswith eachof the consideredclassifiers for thepresent
datasetaremeasured. Then, theclassifier thatperformsbetter inallormost cases isused to justify
thenewlyhypothesized ’uniformpattern’ inRLBPi.e.,RULBP. It is tobenoted thatoneof thekey
parametersofRULBPis thwhosevalue is subjective to thedocument image.Here,different trial runs
52
back to the
book Document Image Processing"
Document Image Processing
- Title
- Document Image Processing
- Authors
- Ergina Kavallieratou
- Laurence Likforman-Sulem
- Editor
- MDPI
- Location
- Basel
- Date
- 2018
- Language
- German
- License
- CC BY-NC-ND 4.0
- ISBN
- 978-3-03897-106-1
- Size
- 17.0 x 24.4 cm
- Pages
- 216
- Keywords
- document image processing, preprocessing, binarizationl, text-line segmentation, handwriting recognition, indic/arabic/asian script, OCR, Video OCR, word spotting, retrieval, document datasets, performance evaluation, document annotation tools
- Category
- Informatik