Handwriting recognition using cohort of LSTM and lexicon verification with extremely large lexicon

Abstract

In this article, a handwriting recognition model whose complexity does not depend on the lexicon size is proposed. It is an alternative to lexicon-driven decoding, based on a lexicon verification process that allows to deal with millions of words, without any time consuming decoding stage. This lexicon verification is included in a cascade framework that uses complementary LSTM RNN classifiers. An original and very efficient method to obtain hundreds of complementary LSTM RNN extracted from a single training, called cohort, is proposed. The proposed approach achieves new state-of-the art performance on the Rimes and IAM datasets, and provides 90% of accuracy on the Rimes dataset when dealing with a gigantic lexicon record of 3 millions of words. The last contribution extends the idea of cohort and lexicon verification in a ROVER combination for handwriting line recognition, and achieves state-of-the-art results on the Rimes dataset.