Large vocabulary off-line handwritten word recognition

Considerable progress has been made in handwriting recognition technology over the last few years. Thus far, handwriting recognition systems have been limited to small-scale and very constrained applications where the number of different words that a system can recognize is the key point for its performance. The capability of dealing with large vocabularies, however, opens up many more applications. In order to translate the gains made by research into large and very-large vocabulary handwriting recognition, it is necessary to further improve the computational efficiency and the accuracy of the current recognition strategies and algorithms. In this thesis we focus on efficient and accurate large vocabulary handwriting recognition. The main challenge is to speedup the recognition process and to improve the recognition accuracy. However, these two aspects are in mutual conflict. It is relatively easy to improve recognition speed while trading away some accuracy. But it is much harder to improve the recognition speed while preserving the accuracy. First, several strategies have been investigated for improving the performance of a baseline recognition system in terms of recognition speed to deal with large and very-large vocabularies. Next, we improve the performance in terms of recognition accuracy while preserving all the original characteristics of the baseline recognition system: omniwriter, unconstrained handwriting, and dynamic lexicons. The main contributions of this thesis are novel search strategies and a novel verification approach that allow us to achieve a 120 speedup and 10% accuracy improvement over a state-of-art baseline recognition system for a very-large vocabulary recognition task (80,000 words). The improvements in speed are obtained by the following techniques: lexical tree search, standard and constrained lexicon-driven level building algorithms, fast two-level decoding algorithm, and a distributed recognition scheme. The recognition accuracy is improved by post-processing the list of the candidate N-best-scoring word hypotheses generated by the baseline recognition system. The list also contains the segmentation of such word hypotheses into characters. A verification module based on a neural network classifier is used to generate a score for each segmented character and in the end, the scores from the baseline recognition system and the verification module are combined to optimize performance. A rejection mechanism is introduced over the combination of the baseline recognition system with the verification module to improve significantly the word recognition rate to about 95% while rejecting 30% of the word hypotheses.