Seite - 137 - in Document Image Processing

J. Imaging 2018,4, 15 3.5.DecodingwithDeepOpticalModels Decoding for bothdeepnet systemswasperformedwithWeighted Finite State Transducers (WFST). Our decoder is based on the CTC-specific implementation proposed by [43] for speech recognition.A“token”WFSTwasdesignedtohandleallpossible label sequencesat the framelevel, soas toallowfor theoccurrenceof theblank labelalongwith therepetitionofnon-blank labels. It can mapasequenceof frame-levelCTC labels toa single character. Asearchgraph isbuiltwith three WFSTs(T,LandG) compiled independentlyandcombinedas follows: S=T◦min(det(L◦G)) (3) T,LandGarethetoken, lexiconandgrammarWFSTsrespectively,whereas◦,detandmindenote composition, determinationandminimization, respectively. Thedeterminationandminimization operationsareneededtocompress thesearchspace,yieldingafasterdecoding. 3.6. EvaluationMetrics Thequalityof theobtainedtranscriptionswasassessedusingtheeditdistance [44]withrespect to thereference text, at thewordandat thecharacter level. TheWordErrorRate (WER) is thisedit distanceat thewordlevelandcanbecalculatedas theminimumnumberofsubstitutions,deletions and insertionsneeded to transform the transcription into the reference, dividedby thenumberof wordsof thereference: WER= s+d+ i n ·100 (4) where s is thenumberof substitutions,d thenumberofdeletions, i thenumberof insertionsandn the totalnumberofwords in thereference. Similarly, this editdistancecanbecalculatedat thecharacter level, giving theCharacterError Rate (CER). In this framework, theCERvalue isespecially interesting, since transcriptionerrorsare usuallycorrectedat thecharacter level. TheOOVWordAccuracyRate (OOVWAR)wasmeasuredas theamountofrecognizedOOVwordsover thetotalamountofOOVwords. Thestatisticalsignificance ofexperimental resultscanbeestimatedbymeansofconfidence intervals.Generally,whencomparing twoexperimental results, it isalways true that if theconfidence intervalsdonotoverlap,wecansay that thedifference is statisticallysignificant [45]. In thiswork, confidence intervalsofprobability95% (α=0.025)werecalculatedbyusingthebootstrappingmethodwith10,000repetitions [46] for these ratemeasures. Finally, as language models are probability distributions over entire sentences or texts, perplexity [47]canbeusedtoevaluate theirperformanceoverareference text. In thiswork,weuse theperplexitypresentedbyacharacterLMover theOOVwords (assequencesofcharacters), toassess thedifferencesbetweentherecognizedandunrecognizedOOVwords. 4. ExperimentalResults In the test experiments, we compared the performance on the test partition of the Rodrigo corpus. Different systemswere compared, the first one based onHMMs, the second one based onRNNandthe thirdoneonCRNN.For the threesystems, experimentswereperformedatword, sub-word,andcharacter levels.Wefirstexplore the influenceof thesizeof theLMcontext (n-gram degree). Then,wedevelop an analysis of thedifference between the structure of recognized and unrecognizedOOVwords. The lastexperimentcompares theresultsobtained in threedifferentcases: openvocabulary, closedvocabularyandwhenusingthevalidationsamples for trainingtheLM. Weobservedthat in the trainingpartitionofRodrigo,usually therearenospacesbetweenwords and punctuationmarks, sowe decided to remove those spaces from the hypotheses offered by theword-basedsystems. Therefore, in theword-basedcases, therecognizedOOVwordscorrespond 137