An Experimental Study of Pruning Techniques in Handwritten Text Recognition Systems

摘要

Handwritten Text Recognition is a problem that has gained attention in the last years mainly due to the interest in the transcription of historical documents. However, the automatic transcription of handwritten documents is not error free and human intervention is typically needed to correct the results of such systems. This interactive scenario demands real-time response. In this paper, we present a study comparing how different pruning techniques affect the performance of two freely available decoding systems, HTK and iATROS. These two systems are based on Hidden Markov Models and n-gram language models. However, while HTK only considers 2-gram language models, iATROS works with n-grams of any order. In this paper, we also carried out a study about how the use of n-grams of size greater than two can enhance results over 2-grams. Experiments are reported with the publicly available ESPOS-ALLES database.