Handwritten Text Recognition Results on the Bentham Collection with Improved Classical N-Gram-HMM methods

Abstract

Handwritten Text Recognition experiments and results are presented on the historical Bentham text image dataset used in the ICFHR-2014 HTRtS competition. The successful segmentation-free holistic framework is adopted, using traditional modelling approaches based on Hidden Markov optical character models (HMM) and an N-gram language model (LM). Departing from the very basic N-gram-HMM baseline system provided in HTRtS, several improvements are made in text image feature extraction and LM and HMM modeling, including more accurate HMM training by means of discriminative training. As a result, we achieve similar recognition accuracy as some of the best performing (single, uncombined) systems based on (recurrent) Neural Networks (NN), using identical training and testing data. The traditional N-gram/HMM framework offers several advantages over modern approaches based on (hybrid, recurrent) NNs. Perhaps the most important is the much faster training of HMMs and the well understood stability of the results of Baum-Welch training. These advantages become very important when dealing with many historical document collections, which are typically huge and entail very high degrees of variability, making it generally difficult to re-use models trained on previous collections.

Tasks

Reproductions