HMM-based Handwritten Word Recognition System by using Singularities Sebastiano Impedovo, Anna Ferrante and Raffaele Modugno Dipartimento di Informatica – Università di Bari – Via Orabona, 4 – 70126 Bari – Italy Centro “Rete Puglia” – Università di Bari – Via G. Petroni, 15/F.1 – 70126 Bari - Italy impedovo@di.uniba.it Abstract This paper presents a new approach for Handwritten Word Recognition based on Hidden Markov Model theory and the sliding window technique. The new approach uses specific singularity markers to support the recognition phase: the Static Marker and the Dynamic Marker. Moreover, different strategies for sliding window step are considered: Regular Step and Progressive Step. Experimental results showing the improvements obtained for basic word lexicon recognition are reported in the paper. 1. Introduction Handwritten recognition is a complex process, in fact during the last thirty-five years a lot of approaches have been proposed and several algorithms have been developed for handwritten digit, character and word recognition in a large variety of application fields [8, 9, 14, 15, 17, 23]. Without any doubt the total information about a grapheme, representing a handwritten trace, includes both the shape and the dynamics of its tracing process but in written words however, the history of this tracing is completely disregarded. On the other hand, for several millennia, the nature of the writing process has ensured knowledge transmission from generation to generation among humans. For this reason, for a long time researchers were not very interested in using algorithms for time dependent processing as theory of Markov chains. However, several attempts to use Markov chains theory for character recognition were made in the ’70 years [10, 11] but based on the poor experimental results achieved no long Markov theory use attracted the scientific community. In fact, after a better understanding of the computational theory and specifically after the successful formulation made by Cooley and Tukey [4] of the FFT algorithms, the Neural Network computation [7] and Rabiner’s idea to use dynamic programming to handle Markov chains of the first order for speech recognition [19], the theory of Hidden Markov Model (HMM) [18] has been used at beginning of the ‘80 years also for off-line Handwritten Word Recognition (HWR) [2, 3, 16]. For this purpose, one of the most interesting technique in using the HMM for HWR, is the artificially recovery of the time dependence of a trace by using a sliding window [25]. Some advancement of the approach with the aim of investigating human mechanisms for handwritten recognition was developed. Furthermore, some experiments to develop very robust algorithms for word recognition have been also proposed in the last years by some researchers [21, 22, 24]. This paper presents a new version of a HWR system, based on the HMM theory and the use of the sliding window technique together new singularity markers and different sliding window step strategies. The algorithms and prototypes have been developed by using an Integrated Development Environment (IDE) software based on a visual programming language. In order to describe the progress and results obtained, this paper reports: in Section 2, shortly the HMM theory and the singularity markers; in Section 3, the recognition approachs based on the sliding window technique and the HMM; in Section 4, an overview of the system prototype with experimental results. Finally, Section 5 reports the conclusions and some suggestions for future improvements of the system. 2. HMM and Singularity Markers This section presents the HMM theory according to the singularity-based approach for HWR and the new strategies based on a marking procedure to implement the recognition algorithms. An HMM is a double stochastic process. The first process is not observable or hidden, the second one produces a sequence of observable symbols according to the probabilistic law that associates the given observed symbol to the hidden state. An HMM is defined as a five-tuple 2009 10th International Conference on Document Analysis and Recognition 978-0-7695-3725-2/09 $25.00 © 2009 IEEE DOI 10.1109/ICDAR.2009.73 783