Handling within-writer variability and between-writer variation in the recognition

Although the performance of current automatic recognition algorithms of on-line handwriting has much improved in recent years, there are still many problems with the actual application of these systems. It appears that the step from academical experiments to real-life use of such algorithms, in, e.g., portable pen computers, is still difficult. What is particularly intriguing is the fact that reported academic (and commercial) recognition rates usually are 10-20% overestimated. When such systems are given the real test, i.e., use by any writer, in a realistic application such as note taking during lectures, their performance drops sharply. One reason lies in the fact that for on-line handwriting, only limited training databases exist. A project is currently running to alleviate this problem: UNIPEN (Guyon & Schomaker, 1994). Although the availability of huge databases for system training and development potentially improves the performance of existing algorithms due to the wider coverage of handwriting shapes, it is very likely that many algorithms are not well fit to handle the case of an infinitely large training set. Neural network-based approaches, but also approaches based on hidden-Markov models both run the risk of satiation, where the system yields an average but incomplete representation of all possible handwriting shapes. Similarly, brute force matching methods run the risk of becoming computationally impractical, when all possible character shapes (allographs) have to be considered. This study is directed at the development of procedures to obtain an insight in the underlying variation of shapes within large quantities of handwriting data from several writers. At this stage, it is useful to make a distinction between two source of variation in handwriting shapes: