Text Segmentation and Grouping in Tactile Graphic Production

by
Matt Renzelmann

Ideas are routinely communicated using a variety of techniques: words, sounds, and images are but a few. Some of these techniques, however, are ineffective if the person being communicated with has a disability, such as blindness. In many science and engineering fields, blind persons face tremendous obstacles in their effort to perceive such forms of communication as diagrams, charts, and graphs. The primary goal of our research is to develop computer software capable of automating the conversion of figures and diagrams in a book to a tactile format that blind persons can readily perceive. Among the challenges posed by this goal is the need to find and remove text from the figures and diagrams for eventual conversion to Braille. It is this challenge on which my work focuses. Since conventional optical character recognition (OCR) software is poorly suited to the task of finding short pieces of text embedded within images, we are exploring new methods based on training and statistics. Because the figures and diagrams in a book are of a similar style, we can exploit this similarity to improve text recognition accuracy. The software we have written is able to recognize a high percentage of the text once provided with a small training set from which to base its predictions. We will discuss the techniques used in our software and make comparisons with alternatives, such as existing commercial OCR software.

Advised by Richard Ladner

CSE 403
Wednesday
April 20, 2005
3:30 - 4:20 pm