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

Abstract

The purpose of on-line handwriting recognition is to automatically recognize characters handwritten with digitizing tablets and pens. For highly accurate character recognition, it is necessary to model character structures as realistically as possible. In this paper, character structures are defined as hierarchical components and their relationships. For instance, a Hangul syllable character has graphemes, strokes and points as its components. An English word has alphabets, strokes and points. Here, strokes are straight or nearly straight traces. Components are hierarchical in that a component consists of its subcomponents and their relationships. Relationships between components are defined as dependencies between their positions. Character structures have not been actively modeled in previous studies. Conventional approaches such as template matching methods, hidden Markov models and time delay neural networks are lack of parameters for explicitly modeling them. They are based on the assumption that local feature inputs in a handwriting input are independent one another. The dissertation proposes a Bayesian network framework which explicitly models character components and their relationships. A character model is composed of grapheme models with inter-grapheme relationships. A grapheme model is composed of stroke models with inter-stroke relationships. Similarly, a stroke model is composed of point models with their relationships. A point model has a 2-D Gaussian distribution for modeling X-Y positions of point instances. Relationships between components are represented by conditional Gaussian distributions. All the models and relationships are probabilistically and graphically represented in a Bayesian network framework. All the parameters are trained from data in the objective of maximum likelihood. The proposed system was trained and evaluated with on-line handwritten Hangul syllables (105 writers, 83,853 characters) and digits (more than 290 writers,...