A Study of Hidden Markov Models for Off-line Recognition of Handwritten Characters

Abstract—In many pattern recognition problems, research has shown that significant improvement in the overall performance may be achieved through a combination of multiple classifiers, rather than a further development of a single classifier [1]. As a process, among others, to reduce the amount of information to be treated by the classifier, one can represent a character image thanks to a sequence of straight lines. Hidden Markov Models (HMMs) are then appropriate to deal with the variability of the length of these sequences, and the one of the relative position of a given segment in the feature vector. Application of HMMs has led us to define an average character for each class, and to consider some new parameters, such as the probability of the length of a sequence, and the probabilities that a sequence ends on, or comes through, a given state. The recognition rate obtained on the NIST3 database [2] is of 84.4 %, using one HMM per class, and of 86 %, using up to 3 HMMs per class. Finally, the association of these HMMs with a neural network trained on the digital images of the characters and offering a recognition rate of 93.1 %, allows to reach an overall recognition rate of 94.1 %.