On-line cursive script recognition using HMM-based letter spotting technique

Abstract

In this thesis, a new framework is proposed for the problem of on-line cursive script recognition. In the proposed framework, the process of word recognition is guided by the psychological model of the human reading process proposed by Taylor and Taylor. It is initially chosen from a lexicon a smaller list of candidate words which have global features equivalent to the unknown input pattern. The reduced lexicon is utilized as linguistic constrains to lattice search. Next, all letter components within the unknown pattern are detected using a letter spotting technique utilizing hidden Markov models (HMM), which has been successfully applied in speech recognition and language modeling problems. A letter hypothesis lattice is generated as a result of letter spotting. Then an island-driven search technique is applied to find the optimal path on the letter hypothesis lattice. In case that lattice search fails to find a complete path covering the entire input data, an error-tolerant word matching procedure is followed to rank all candidate words in the reduced lexicon. The proposed framework can be easily incorporated with other recognition methods such as neural network models. For instance, in order to increase the discrimination power of letter models we use neural network models in the place of HMMs. Thus, we can find on a letter hypothesis lattice the optimal path that maximizes a global criterion with respect to HMM-based latter segmentation and neural network model-based letter classification. This hybrid framework provides a nice recognition scheme of integrating the temporal structure of HMMs with the discrimination power of neural networks. It is shown by an experiment that the proposed framework is promising for recognizing English cursive words.