Ben Holm

Abstract

A revision of Recognition Strategy Language (RSL), a domain-specific language for pattern recognition algorithm development, is in development. This language provides several tools for pattern recognition algorithm implementation and analysis, including composition of operations and a detailed history of those operations and their results. This research focuses on that history and shows that for some problems it provides an improvement over traditional methods of gathering information. When designing a pattern recognition algorithm, bookkeeping code in the form of copious logging and tracing code must be written and analyzed in order to test the effectiveness of procedures and parameters. The amount of data grows when dealing with video streams; new organization and searching tools need to be designed in order to manage the large volume of data. General purpose languages have techniques like Aspect Oriented Programming intended to address this problem, but a general approach is limited because it does not provide tools that are useful to only one problem domain. By incorporating support for this bookkeeping work directly into the language, RSL provides an improvement over the general approach in both development time and ability to evaluate the algorithm being designed for some problems. The utility of RSL is tested by evaluating the implementation process of a computer vision algorithm for recognizing American Sign Language (ASL). RSL history is examined in terms of its use in the development and evaluation stages of the algorithm, and the usefulness of the history is stated based on the benefit seen at each stage. RSL is found to be valuable for a portion of the algorithm involving distinct steps that provide opportunity for comparison. RSL was less beneficial for the dynamic programming portion of the algorithm. Compromises were made for performance reasons while implementing the dynamic programming solution and the inspection at every step of what amounts to a brute-force search was less informative. We suggest that this investigation could be continued by testing with a larger data set and by comparing this ASL recognition algorithm with another.

Library of Congress Subject Headings

Computer vision--Data processing; Pattern recognition systems; American Sign Language--Data processing

Publication Date

2011

Document Type

Thesis

Department, Program, or Center

Computer Science (GCCIS)

Advisor

Fluet, Matthew

Advisor/Committee Member

Zanibbi, Richard

Comments

Note: imported from RIT’s Digital Media Library running on DSpace to RIT Scholar Works. Physical copy available through RIT's The Wallace Library at: TA1634 .H65 2011

Campus

RIT – Main Campus

Download

Share

COinS