NSHP-HMM Based on Conditional Zone Observation Probabilities for Off-Line Handwriting Recognition

Abstract

This work aims at improving the recognition accuracy of the two-dimensional stochastic model NSHP-HMM. The key feature of the modified model is the use of the NSHP Markov random field to describe the contextual information at a zone level rather than a pixel level. Therefore, the use of highlevel features extracted directly on the gray-level zones is permitted, unlike what is done in a recognition based on classical NSHP-HMM where the model, mandatory, operates at a pixel level on normalized binary images. First experiments on handwritten digit recognition show that the proposed model outperforms the classical NSHP-HMM.