Abstract: This paper presents a new strategy for isolating handwritten characters in cursive words without making an explicit a priori segmentation of the script, and without imposing any lexical or other linguistic constraints. Furthermore, this approach can be completely trained using data sets of isolated characters only.
The main idea behind this strategy is to have a window of attention moving around in the cursive word, searching for instances of known characters. If one assumes that the current window contains some significant part of a character, then the problem is to translate and scale the window of attention in such a way that it converges to the bounding box of that character. This process is implemented using both a detector network and a set of locator networks. The detector network is responsible for recognizing whole characters of any class and thus for stopping the iterative process, whereas a locator network is assigned the task of recognizing the crucial parts of a given character class and producing the corresponding transformation parameters for the window. The feasibility of this process is shown through experiments using the UNIPEN database of on-line scripts.