Can scanning n-tuple classifiers be improved by pre-transforming training data?

A new method of applying n-tuple recognition techniques to handwritten OCR has recently been reported, which involves scanning an n-tuple classifier over a chain-code of the image. In scanning n-tuple systems, the traditional advantages of n-tuple recognition i.e. training and recognition speed are retained, while offering superior recognition accuracy, as demonstrated by results on three widely used data sets. Furthermore, the scanning n-tuple systems are less liable to saturation than conventional n-tuple classifiers. One effect of this is that even for large datasets, the training set accuracy remains extremely close to 100%. This paper explores the idea of expanding the training set artificially by pre-transforming the images in various ways, the aim being to improve recognition accuracy on unseen data.