Jigsawing : A Method to Create Virtual Examples in OCR data

In this theoretical note we propose the use of a suffix tree on square matrices for compact representation of a set of training patterns. We show how a test pattern can be generated by jigsawing various regions from different training patterns. This in turn leads us naturally to a compact data dependent representation of a test pattern which we call the description tree. We envisage the use of the description tree in a variety of applications including nearest neighbor classifiers, data dependent distance norms, kernel methods and syntactic pattern recognition. We provide statistical learning theory based arguments to show that our method generates valid virtual examples and hence will lead to better classification accuracy.