Abstract
Software for recognition of handwriting has been available for
several decades now and research on the subject have produced
several different strategies for producing competitive recognition
accuracies, especially in the case of isolated single characters.
The problem of recognizing samples of handwriting with arbitrary
connections between constituent characters (emph{unconstrained
handwriting}) adds considerable complexity in form of the
segmentation problem. In other words a recognition system, not
constrained to the isolated single character case, needs to be able
to recognize where in the sample one letter ends and another begins.
In the research community and probably also in commercial systems
the most common technique for recognizing unconstrained handwriting
compromise Neural Networks for partial character matching along with
Hidden Markov Modeling for combining partial results to string
hypothesis. Neural Networks are often favored by the research
community since the recognition functions are more or less
automatically inferred from a training set of handwritten samples.
From a commercial perspective a downside to this property is the
lack of control, since there is no explicit information on the types
of samples that can be correctly recognized by the system. In a
template based system, each style of writing a particular character
is explicitly modeled, and thus provides some intuition regarding
the types of errors (confusions) that the system is prone to make.
Most template based recognition methods today only work for the
isolated single character recognition problem and extensions to
unconstrained recognition is usually not straightforward. This
thesis presents a step-by-step recipe for producing a template based
recognition system which extends naturally to unconstrained
handwriting recognition through simple graph techniques. A system
based on this construction has been implemented and tested for the
difficult case of unconstrained online Arabic handwriting
recognition with good results.
several decades now and research on the subject have produced
several different strategies for producing competitive recognition
accuracies, especially in the case of isolated single characters.
The problem of recognizing samples of handwriting with arbitrary
connections between constituent characters (emph{unconstrained
handwriting}) adds considerable complexity in form of the
segmentation problem. In other words a recognition system, not
constrained to the isolated single character case, needs to be able
to recognize where in the sample one letter ends and another begins.
In the research community and probably also in commercial systems
the most common technique for recognizing unconstrained handwriting
compromise Neural Networks for partial character matching along with
Hidden Markov Modeling for combining partial results to string
hypothesis. Neural Networks are often favored by the research
community since the recognition functions are more or less
automatically inferred from a training set of handwritten samples.
From a commercial perspective a downside to this property is the
lack of control, since there is no explicit information on the types
of samples that can be correctly recognized by the system. In a
template based system, each style of writing a particular character
is explicitly modeled, and thus provides some intuition regarding
the types of errors (confusions) that the system is prone to make.
Most template based recognition methods today only work for the
isolated single character recognition problem and extensions to
unconstrained recognition is usually not straightforward. This
thesis presents a step-by-step recipe for producing a template based
recognition system which extends naturally to unconstrained
handwriting recognition through simple graph techniques. A system
based on this construction has been implemented and tested for the
difficult case of unconstrained online Arabic handwriting
recognition with good results.
| Original language | English |
|---|---|
| Qualification | Doctor |
| Awarding Institution |
|
| Supervisors/Advisors |
|
| Award date | 2008 May 30 |
| Publication status | Published - 2008 |
Bibliographical note
Defence detailsDate: 2008-05-30
Time: 13:15
Place: Lecture Hall MH:C, Centre for Mathematical Sciences, Sölvegatan 18, Lund university, Faculty of Engineering
External reviewer(s)
Name: Srihari, Sargur
Title: Prof
Affiliation: CEDAR, NY, USA
---
Subject classification (UKÄ)
- Computer Vision and Robotics (Autonomous Systems)
- Mathematics
Free keywords
- on-line handwriting recognition segmentation dynamic programming clustering discriminative clustering