Extracting On - Line Data from O - Line

A heuristic method for simulating time dependent data from oo-line character input is proposed. The method involves extracting strokes from a thinned bitmap and then ordering them along with simulated pen-up strokes to achieve a pseudo-time-dependent representation. Features used for input to time-delay neural nets 1] and 2] are then extracted. The algorithm is tested on fortyinputs and the feature matrices extracted are compared statisticallyagainst each other, showing promising results. Historical Background The problem of computer recognition of handwritten characters has vexed researchers for years, providing an ever challenging eld to explore. The applications of such research seem endless and would include systems in the areas of postal code identiication, check processing, and more natural human-computer interfaces 6]. Traditionally, the eld can be divided into two domains: on-line processing, where information regarding time-sequencing of points used to draw the character is known, and oo-line processing, where no such information is available-just a picture of the character. The two domains of oo-line and on-line character recognition have in the past been separate due to the diierence in problem definition .This has lead to a number of diierent approaches to solving each, with little overlap 3]. On-line recognition is an easier problem to solve due to the presence of time related data and the absence of noise and has therefore lead to greater success rates 4], 5]. It would make sense, therefore, to map the oo-line domain onto the on-line domain. Then solving the oo-line problem would be the same as the on-line, assuming the mapping is accurate and appropriate. In other words, it would be desirable to be able to extract time dependent data from a static oo-line image. Little has been done in this area, perhaps because some oo-line techniques have lead to acceptable recognition rates. It has been suggested that such a mapping be devised 4], although it appears that only preliminary work has been done 6]. It has also been suggested that the inverse mapping (from oo-line to on-line data) be attempted 7], so a union of the two elds seems to be of merit. Using a time-delay neural net, Guyon 1] achieved an impressive 96.6% recognition rate for on-line digits and uppercase handwritten characters. These re