Ensemble methods for offline handwritten text line recognition

This thesis investigates ensemble methods for offline recognition of English handwritten text lines. Multiple recognisers are automatically generated from a single base recognition system. Combining the output of these multiple recognisers provides the final ensemble result. The underlying recognisers are based on hidden Markov models. One model is built for each character. Based on the lexicon, word models are derived by concatenating character models. A statistical language model is used to build text line models by preferring more likely word sequences over unlikely word sequences. A postprocessing step calculates confidence values for each recognised word. Ensembles of recognisers are generated based on variation of the training data, the features, and the system architecture. Because the output of a handwritten text line recogniser is a sequence of words, most existing combination methods cannot be applied directly. The combination has to be performed in two steps. First, the word sequences are synchronised by a string alignment procedure. Second, a decision strategy derives the combination result for each segment of the alignment. For this purpose, confidence-based voting, a statistical decision method, and a decision method that includes language model information are used. The experimental evaluation on a large set of images of handwritten text lines indicates that the proposed ensemble methods can significantly increase the performance of an offline handwritten text line recognition system.