Bayesian network modeling of strokes and their relationships for on-line handwriting recognition

Abstract

It is important to model strokes and their relationships for on-line handwriting recognition, because they reflect character structures. We propose to model them explicitly and statistically with Bayesian networks. A character is modeled with stroke models and their relationships. Strokes, parts of handwriting traces that are approximately linear, are modeled with a set of point models and their relationships. Points are modeled with conditional probability tables and distributions for pen status and X, Y positions in the 2-D space, given the information of related points. A Bayesian network is adopted to represent a character model, whose nodes correspond to point models and arcs their dependencies. The proposed system was tested on the recognition of on-line handwritten digits. It showed higher recognition rates than the HMM based recognizer with chaincode features and was comparable to other published systems.