Browse Prior Art Database

Handwriting Recognition System That Uses Both Static and Dynamic Features

IP.com Disclosure Number: IPCOM000021601D
Original Publication Date: 2004-Jan-26
Included in the Prior Art Database: 2004-Jan-26
Document File: 2 page(s) / 45K

Publishing Venue

IBM

Abstract

This invention describes a system which improves the accuracy of machine recognition of dynamic handwriting. The system generates static handwriting data from dynamic handwriting data, generates a machine recognition systems for each data type and combines the two recognition systems to improve the machine recognition of handwritten words.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 51% of the total text.

Page 1 of 2

Handwriting Recognition System That Uses Both Static and Dynamic Features

Typical handwriting recognition systems can be divided into those which use dynamic data (e.g., personal digital assistants and cellphones) and those which use only static data (e.g., document scanners). When recognition is performed using static data, only the image of the handwritten data is used. Using the image, it is possible to know the position of the strokes, but not their temporal order/direction in which they were written. Image data contain ambiguities when letters overlap in the horizontal or vertical directions. These ambiguities make segmentation from sentence to word and from word to character difficult and can result in decreased accuracy. When the recognition is performed using dynamic data, temporal information is available and the strokes can be ordered not only by their position, but also following the time they were produced. This information allows dynamic data to be used to segment overlapping words and characters; thus simplifying the segmentation process. However, dynamic data may also included information irrelevant to recognition (e.g., whether an "O" is written clockwise or counterclockwise) which can increase the complexity of the modeling task and can lead to recognition errors. Static data do not suffer from this problem. Static and dynamic data can thus complement one another.

The proposed system is a combination of two subsystems: a dynamic recognizer and a static recognizer. The dynamic data is recognized by the dynamic system and passed, through a conversion process, to the static system. Both static and dynamic systems convert the data into a sequence of observation vectors and recognize them using Hidden Markov Models (HMM) or similar machine learning algorithms known to those skilled in the art. The system ouputs can be converted into posterior probabilities if desired. The resulting outputs are merged using methods know to those skilled in the art. One possible embodiment of the merging is to use a weighted combination. Since many dynamic and static recognition system errors are uncorrelated, combining the static and dynamic recognizers can improve recognition accuracy.

The first step in the static recognizer is the conversion of the dynamic data from dynamic to static. The dynamic data is a time-orde...