Reconnaissance de l’écriture manuscrite avec des réseaux récurrents
Luc Mioulet · HAL (Le Centre pour la Communication Scientifique Directe) · 2015
Mass digitization of paper documents requires highly efficient optical cha-racter recognition systems. Digital versions of paper documents enable the useof search engines through keyword dectection or the extraction of high levelinformation (e.g. : titles, author, dates). Unfortunately writing recognition sys-tems and especially handwriting recognition systems are still far from havingsimilar performance to that of a human being on the most difficult documents.This industrial PhD (CIFRE) between Airbus DS and the LITIS, that tookplace within the MAURDOR project time frame, aims to seek out and improvethe state of the art systems for handwriting recognition.We compare different systems for handwriting recognition. Our compa-risons include various feature sets as well as various dynamic classifiers : i)Hidden Markov Models, ii) hybrid neural network/HMM, iii) hybrid recurrentnetwork Bidirectional Long Short Term Memory - Connectionist TemporalClassification (BLSTM-CTC)/MMC, iv) a hybrid Conditional Random Fields(CRF)/HMM. We compared these results within the framework of the WR2task of the ICDAR 2009 competition, namely a word recognition task usinga 1600 word lexicon. Our results rank the BLSTM-CTC/HMM system as themost performant, as well as clearly showing that BLSTM-CTCs trained ondifferent features are complementary.Our second contribution aims at using this complementary. We explorevarious combination strategies that take place at different levels of the BLSTM-CTC architecture : low level (early fusion), mid level (within the network),high level (late integration). Here again we measure the performances of theWR2 task of the ICDAR 2009 competition. Overall our results show thatour different combination strategies improve on the single feature systems,moreover our best combination results are close to that of the state of theart system on the same task. On top of that we have observed that some ofour combinations are more adapted for systems using a lexicon to correct amistake, while other are better suited for systems with no lexicon.Our third contribution is focused on tasks related to handwriting recognition. We present two systems, one designed for language recognition, theother one for keyword detection, either from a text query or an image query.For these two tasks our systems stand out from the literature since they usea handwriting recognition step. Indeed most literature systems focus on extracting image features for classification or comparison, wich does not seemappropriate given the tasks. Our systems use a handwriting recognition stepfollowed either by a language detection step or a word detection step, depending on the application.