Page - 142 - in Document Image Processing

J. Imaging 2018,4, 15 4.4. Studyof theContextSize InfluenceUsingDeepOpticalModels This last part of the experimentation studies the influence of the different language units and the context size of the languagemodel, on theHTR systembased ondeep neural networks (seeSections3.4.2and3.4.3). 4.4.1. Results forDeepModelsBasedonRecurrentNeuralNetworkswithBLSTMs InFigure 14, the recognition results obtained for theword-basedRNNsystemarepresented. As explained before, in this case, the recognized OOV words correspond to words attached to punctuationmarks,whichwerecorrectlyrecognizedafter removingthespacebetweenthem(see the examplepresented inFigureA2).Comparedwith theword-basedHMMsystem, theobtainedresults are significantlyworse in terms ofWER; however, in terms of CER andOOVword recognition accuracy, theobtainedresultsaresignificantlybetter. Concretely, thebest resultwasobtainedbyusing a two-gramLM,anditpresentsaWERequal to52.5%±0.8,aCERequal to17.2%±0.3andanOOV WARequal to16.3%±0.9. Figure15showstheresultsobtainedusingsub-wordn-gramLM.Ascanbeobserved, theWFST approachhasnocontext informationabout theseparationbetweenwordswhensub-wordunigrams LMareused; therefore, it isunable toreconstructwordscorrectly inspiteofobtainingagoodCER. Wewill see thiseffect inthenextexperimentswiththesub-wordandcharacter-baseddeepnetsystems. Inthiscase, thebestresultwasobtainedwithafive-gramlanguagemodel (aWERequal to38.6%±0.5, aCERequal to17.3%±0.3andanOOVWARequal to27.4%±1.1). 10% 20% 30% 40% 50% 60% 1 2 3 4 5 6 WER=52.5% CER=17.2% OOV WAR=16.3% n-gram size Word Error Rate Character Error Rate OOV Word Accuracy Rate Figure14.ResultsobtainedbytheRNNword-basedsystemusingn-gramlanguagemodels. 0% 20% 40% 60% 80% 100% 1 2 3 4 5 6 WER=38.6% CER=17.3% OOV WAR=27.4% n-gram size Word Error Rate Character Error Rate OOV Word Accuracy Rate Figure15.ResultsobtainedbytheRNNsub-word-basedsystemusingn-gramlanguagemodels. 142