The efficiency of the NSHPZ-HMM: theoretical and practical study

Abstract

In this paper, we propose a novel HMM-based 2-D recognition engine, namely the NSHPZ-HMM. Like the reference model (the NSHP-HMM), the proposed classifier brings the efficient training and decoding algorithms of 1-D HMM to the 2-D modeling of spatial data. Furthermore, in contrast to the reference model which suffers from the short 2-D context limitation, our model uses the NSHP Markov random field to describe the contextual information at a ’zone’ level rather than a ’pixel’ level; the goal is to extend the context in order to give a better modeling of the spatial property of an image. Therefore, the use of high-level features extracted directly on the gray-level or color zones is possible, unlike what is done in a recognition based on classical NSHP-HMM, where the model, mandatorily, operates at a pixel level on normalized binary images; consequently, the applicability of our model is more general compared to the classical NSHP-HMM. Throughout this paper, we demonstrate the efficiency of the proposed approach at two stages. Firstly, in the theoretical study, we show the advantage of our model over other HMM-based 2-D classifiers; this part constitutes by itself, to our knowledge, the first complete overview of 2-D recognition approaches. Secondly, the experimental evaluation performed on recognition of handwritten digits/words provides the effectiveness of the NSHPZ-HMM against all other HMM-based 2-D recognizers and shows a good potential for other image recognition applications.