Exploration in on-line handwritten character recognition.
Item
-
Title
-
Exploration in on-line handwritten character recognition.
-
Identifier
-
AAI9405556
-
identifier
-
9405556
-
Creator
-
Matic, Nada Petar.
-
Contributor
-
Adviser: Nenad Marinovich
-
Date
-
1993
-
Language
-
English
-
Publisher
-
City University of New York.
-
Subject
-
Engineering, Electronics and Electrical | Computer Science | Artificial Intelligence
-
Abstract
-
In this thesis we address the problem of learning and recognition of the handwritten isolated characters. The "core" of the recognition system is Time Delay Neural Network (TDNN) which incorporates feature extractor and classifier in a single trainable module. Our experiments were carried out on a database of handwritten characters, entered on a touch terminal.;Specifically, the objective of this research was to improve the recognition rate by focusing on the most informative patterns. They are examples of rare writing styles that are typically under-represented in the training database. We have used two different approaches in order to pay special attention to these patterns.;One approach is applicable to the writer independent task and consists of developing a systematic computer-aided methodology for cleaning the large training database. Using this method, non-informative (e.g. meaningless and mislabeled) patterns are removed from the training database. The rare and unusual patterns that remain in the tail of the distribution after the meaningless or mislabeled patterns have been removed are then emphasized using special training procedure. By combining cleaning and emphasizing training scheme we have reduced the generalization error by a factor of two.;Second approach is to introduce writer adaptation to allow a specific user to provide, on-line, examples of rare character styles that do not exist in the database. To do that, we use the writer independent neural network without its last layer as a preprocessor to the Optimal Hyperplane Classifier. Through a process of adaptation, we retrain the Optimal Hyperplanes with the set of most informative patterns of the default database, augmented by the examples provided on-line by a specific user. Without degradation in speed, the average error rate after adaptation is 1% to 2% for most writers.
-
Type
-
dissertation
-
Source
-
PQT Legacy CUNY.xlsx
-
degree
-
Ph.D.
