Seite - 139 - in Document Image Processing

J. Imaging 2018,4, 15 0% 10% 20% 30% 40% 50% 60% 70% 80% 1 2 3 4 5 6 WER=43.2% CER=20.0% OOV WAR=9.3% n-gram size Word Error Rate Character Error Rate OOV Word Accuracy Rate Figure8.Resultsobtainedbydecodingat theHMMsub-word levelbyusingn-gramlanguagemodels withsizen={1,. . . ,6}. 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 WER=39.8% CER=17.6% OOV WAR=18.3% n-gram size Word Error Rate Character Error Rate OOV Word Accuracy Rate Figure9.Resultsobtainedbydecodingat theHMMcharacter levelbyusingn-gramlanguagemodels withsizen={1,. . . ,15}. Table 2. Overall best results on the Rodrigo test set in terms ofWER, CER andOOVWAR for theHMMsystem. Measure Word Sub-Word Character3-gram 4-gram 10-gram WER 43.9%±0.5 43.2%±0.5 39.8%±0.5 CER 21.2%±0.3 20.0%±0.3 17.6%±0.3 OOVWAR 2.3%±0.3 9.3%±0.7 18.3%±0.9 4.2. Studyof theRelationbetween theStructureof theOOVWordsandtheTrainingWords Thecharacter-basedapproachisabletorecognizesomeOOVwordsgiventhatthecharacter-based LMlearns thestructureof thewordscontainedinthe trainingset. Inorder toverify thishypothesis, wemeasuredtheperplexitypresentedbythebestcharacter-basedLM(10-gram)fordecodingeachone of the4918OOVwordsas theircorrespondingcharactersequences. Figure10presents theobtained perplexityperOOVwordseparated into twodistributions, recognizedandunrecognizedOOVwords. Table3summarizes themainfeaturesof thesedistributions.Asexpected, therecognizedOOVwords present lowerperplexity thantheunrecognizedOOVwords. Theoverlapofbothdistributionsmakes us thinkthat there is still roomfor improvementgiventhatmoreOOVwordscouldberecognized. 139