Seite - 144 - in Document Image Processing

J. Imaging 2018,4, 15 Regarding the recognition ofOOVwords, the sub-word approach allowed correctly recognizing 42.4%±1.5of theOOVwords. 5% 10% 15% 20% 25% 30% 1 2 3 4 5 6 WER=17.9% CER=4% OOV WAR=21.5% n-gram size Word Error Rate Character Error Rate OOV Word Accuracy Rate Figure17.Resultsobtainedby theCRNNword-basedsystemusingn-gramlanguagemodelswith sizen={1,. . . ,6}. 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 1 2 3 4 5 6 WER=14.8% CER=3.4% OOV WAR=42.4% n-gram size Word Error Rate Character Error Rate OOV Word Accuracy Rate Figure18.ResultsobtainedbytheCRNNsub-word-basedsystemusingn-gramlanguagemodelswith sizen={1,. . . ,6}. Figure19presents theresultsobtainedwith theCRNNsystemusingcharactern-gramLM.As in thepreviouscharacter-basedexperiments, similar resultsareobtainedforn≥6,andtheoverallbest resultwasobtainedwitha10-gramcharacter languagemodel (aWERequal to14.0%±0.3andaCER equal to3.0%±0.1). RegardingtherecognitionofOOVwords, thisapproachwasable torecognize correctly 69.2%±1.1 of theOOVwords using no external resource or dictionary, but a character languagemodelonly. Table 5 presents a summary of the obtained best results for the test experiments for the CRNN system. As can be observed, the use of deep optical models allows one to obtain astatistically-significantrelativeimprovementof59.2%overtheHMMsystem(43.9%±0.5) intermsof WERand81.1%statistically-significant relative improvementover theHMMsystemintermsofCER. RegardingOOVwords,21.5%ofOOVwords,whichcorrespondtowords followedbypunctuation marks,arewell recognized. It shouldbenotedthat theseresultsarealsosignificantlybetter thanthose obtainedbytheHMMsystemintheclosedvocabularyexperiments (Figures11–13). Theuseofsub-wordunitsperformsbetter thanusingwords. In thiscase, theuseofa four-gram LMtrainedwithhyphenatedwordsallowedobtaining statistically-significant improvementsover theuseof a three-gramLMof fullwords. However, theoverall best results areobtainedbyusing 144