Machine learning applied to recognising hand-written Japanese

Abstract

This paper considers a machine learning technique applied to handwriting recognition for both Japanese Hiragana characters and English capital letters. Machine induction is used to obtain decision trees from examples. The main advantage that rule induction offers over other handwriting recognition techniques is the ability to learn an algorithm for each user of a computer system and thus to offer a unique recognition algorithm for each user. For this study, actual examples of handwriting are analysed in order to produce a simulation program which generates many thousands of examples. The whole process of simulation, through induction of a decision tree to final testing on the example set is automated. Provisional results are presented for simulation and recognition of Japanese Hiragana characters and for English capital letters. The decision trees induced are shown and a critical analysis of the approach is given. Provisional results suggest that the choice of primitives is critical, and the induced algorithm requires only selective evaluation of these primitives and higher order derivatives.