Handwritten word recognition using hidden Markov models
Mou-Yen Chen · 1993
Research in text recognition has long been focused on the recognition of machine-printed characters. Because of large variation involved in human handwriting, the recognition problem is very difficult. The Hidden Markov Model (HMM) has been widely and successfully used in speech processing and recognition. The recognition of handwritten words has many similarities to that of speech. They both involve the processing of noisy language symbol strings with ambiguous boundaries and considerable variations in symbol appearance. In this dissertation, we have proposed various types of HMM's for use in the recognition of totally unconstrained handwritten words (i.e. the words might by discretely printed, cursive, or a combination of both). In applying HMM's to pattern recognition problems, the first consideration is how to represent the evolutionary pattern as a sequence of observations such that the HMM is applicable. The solution is quite straightforward for speech because of its strong temporal constraints. For handwriting, especially in off-line recognition systems, it is not trivial. Thus, a good segmentation algorithm is needed to obtain the observation sequence for the HMM's. Exclusively for this purpose, we have developed, and detailed in this dissertation, a sub-character segmentation algorithm based on mathematical morphology. Another important reason for the HMM's success in speech recognition is its effective representation of speech in terms of LPC features. Unfortunately, no satisfactory representation exists for handwriting signals, and finding such a representation is an open research problem. In the approaches proposed here, based on empirical evidence, we have carefully selected a feature set which includes moment features, geometrical and topological features, and local and global pixel distribution features. Finally, an important issue for applying HMM's is to select a good unit of handwriting to be modeled by the HMM. As in speech, words are the natural units for handwriting. Another alternative is the character, which is somewhat like the phoneme of speech but with a more distinct meaning. We have designed a number of systems for handwriting recognition using different HMM units and two different implementation strategies: the model discriminant HMM and the path discriminant HMM. The success of these systems is demonstrated by means of experiments using real-world data. Finally, the comparative advantages and disadvantages of different systems are discussed.