Recognition of on-line cursive script
Some of the main problems associated with recognition of on-line cursive script are how to deal with extremely large variability of letter shapes and how to deal with the segmentation problem. The variability in letter shapes is partially a result of different writing styles but even for the same writer the same letter is often written differently depending on the position of the letter within a word.To get a sense of the variability of letter shapes check out the following examples from our database (obtained from David Rumelhart). If you were not able to read all the words, here is the list (ordered from left to right, top to down): animal, foolish, river, account, art, and sleep.
One of the reasons we started this project is because of the segmentation problem. Although extremely important in many areas, including computer vision and sequence analysis (from DNA analysis to speech recognition), the segmentation problem remains largely unsolved. Why is that? Probably because it is one of those "chicken-and-egg" problems. It is very hard to get a handle of those problems. In order to segment a word it helps to know the letters but in order to recognize the letters it helps to know the word. Recognition of on-line script seemed like a nice starting point since patterns are one-dimensional but at the same time sufficiently complex and challenging. In the following, I will briefly describe two models that we have developed for recognition of cursive script. The main advantage of these two models over the state-of-the-art model for cursive script recognition, the Hidden Markov Model (HMM), is that they can handle duration modeling without an increase in computational complexity.