Chinese Journal of Computers   Full Text
Title	Cascaded HMM Training Algorithm for Continuous Character Recognition
Authors	ZHAO Wei LIU Jia-Feng TANG Xiang-Long WU Rui
Address	(School of Computer, Harbin Institute of Technology, Harbin 150001)
Year	2007
Issue	No.12(2142—2150)
Abstract &   Background	Abstract Continuous handwritten recognition, including the recognition of sentences, words and character sequences is a new and important branch in character recognition. One of the key techniques in Continuous Handwritten Character Recognition is how to model every word entry in lexicon and every sentence that made up by the words. Due to the HMM′s characteristic of time sequence modeling capability, a Cascaded Hidden Markov Models (Cascaded HMM), which defines model connection probability and state transition probability between HMMs, is proposed in this study. By modifying the Baum-Welch algorithm reestimation formula, Cascaded HMM reestimates connection parameter of character HMMs. The description of Cascaded Baum Welch Algorithm and Cascaded Viterbi Algorithm are given. The cascaded idea of recognizing continuous character is with the strategy of free-segmentation and dynamic programming. Meanwhile, Cascaded HMMs do not model every item listed in lexicon, but combine character models as continuous text model. The cascaded HMM method could accurately describe the shape variability between adjacent characters in handwritten curve. In handwritten English word recognition task, test result shows that the cascaded model is prior to the baseline system. The method offers strong support for continuous recognition technology. keywords continuous handwritten character recognition; hidden Markov models; cascaded model; state transition between models; cascaded training background Since the 1990s, interests of Hidden Markov Model (HMM) for handwritten character recognition have been growing significantly. As an effective combination of Markov chain and stochastic process, HMM could depict random signal using state transition and statistic mapping relation between state and observation. While in continue character recognition, it is difficult to establish an independent statistical model for each time sequence, or we say handwritten word, which sampled from tablet, because of the variability and illegibility of free handwritten text. How to represent a word model by using as less character HMM as possible is one of the hotspot in this field. Researchers have been making sustained attempt at novel approaches. In this paper, a continuous stochastic model, named as cascaded HMM, is proposed. It could generalize an integrated word model for certain a free handwritten word by using single character HMMs. The key of the method is the probability of shifting between character models and the cascaded Baum-Welch Algorithm that are explained elaborately in the paper. By using Baum-Welch and cascaded Baum-Welch training, only 52 character HMMs are obtained. These HMMs will be connected as word level cascaded HMM and used in decoding free handwritten words. Without label any of the word, the experiment results show effectiveness of the cascaded algorithm and higher performance in online English word recognition. This group has been working in off-line or on-line character recognition research for many years. Their interests have been in both pen-based device design and recognition strategy and algorithm design. Till now, they had gained many research fruit and reward in this area. By the year of 2006, they had published more than 100 papers in the international proceedings and journals on this topic.