Seite - 202 - in Document Image Processing
Bild der Seite - 202 -
Text der Seite - 202 -
J. Imaging 2018,4, 32
recognition protocols in terms ofCRR,WRRandLRRmetrics. The SID-OCR systemhas shown
superiority inallprotocols. ThebestaccuraciesareachievedontheTunisiaNat1channel subset (p6.3)
with0.94asaCRRand0.62asaLRR.TheIWATAsystemperformswell forallSDprotocolsespecially
for theCRR/WRRmetrics.However itscurrentversionis incompatiblewithHDresolution. Theresult
showsthatoursystemhas lowrecognitionratewhenfacingdifferent textpatternsandresolutions,
i.e.,globalProtocol9. Basedonourknowledgeabout theshapesofArabiccharacters,wedivide the
causesoferrors into twoclasses: charactersimilarityandinsufficientsamplesofpunctuation,digits
and symbols. Severalmeasures canbe taken tominimize the character error rate, for instanceby
integrating languagemodelsordropoutmechanism.
Table7.Performanceof textdetectionsystemsevaluatedonthe test setofAcTiV-D.
Protocol System Precision Recall Fmeasure
1 LADI[46] 0.86 0.84 0.85
SysA[14] 0.77 0.76 0.76
Gaddo[52] 0.52 0.49 0.51
4.1 LADI[46] 0.74 0.76 0.75
SysA[14] 0.69 0.6 0.64
Gaddo[52] 0.47 0.61 0.54
4.2 LADI[46] 0.8 0.75 0.77
SysA[14] 0.66 0.55 0.6
Gaddo[52] 0.41 0.5 0.45
4.3 LADI[46] 0.85 0.82 0.83
SysA[14] 0.68 0.71 0.69
Gaddo[52] 0.34 0.49 0.41
4.4 LADI[46] 0.71 0.76 0.73
SysA[14] 0.5 0.49 0.49
Gaddo[52] - - -
Table8.Performanceof therecognitionsystemsevaluatedonthe test setofAcTiV-R.
Protocol System CRR WRR LRR
3 SIDOCR[51] 0.90 0.71
0.51IWATA[53]
- - -
6.1 SIDOCR[51] 0.89 0.70
0.51IWATA[53]
0.88 0.67 0.46
6.2 SIDOCR[51] 0.94 0.68
0.41IWATA[53]
0.9 0.68 0.39
6.3 SIDOCR[51] 0.94 0.81
0.62IWATA[53]
0.94 0.77 0.56
6.4 SIDOCR[51] 0.93 0.73
0.52IWATA[53]
0.9 0.73 0.48
9 SIDOCR[51] 0.73 0.58
0.32IWATA[53]
- - -
5.3.2. TrainingwithAcTiV2.0
Toexamine theeffectof increasing thenumberof trainingsamplesontheaccuracyofour text
detector, we conduct the same experiment of Protocol 6.1, in Table 7, using training-set2, which
includes roughly thedoubleof samples (600 frames) than training-set1 (seeTable4). Weobserved
that thedetectionratesofour textdetectorhavebeen increasedasexpected. Specifically, the recall
increasesby2%andtheprecisionincreasesby5%.Thiscanbeexplainedbythe increase inthenumber
202
zurück zum
Buch Document Image Processing"
Document Image Processing
- Titel
- Document Image Processing
- Autoren
- Ergina Kavallieratou
- Laurence Likforman-Sulem
- Herausgeber
- MDPI
- Ort
- Basel
- Datum
- 2018
- Sprache
- deutsch
- Lizenz
- CC BY-NC-ND 4.0
- ISBN
- 978-3-03897-106-1
- Abmessungen
- 17.0 x 24.4 cm
- Seiten
- 216
- Schlagwörter
- 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
- Kategorie
- Informatik