University of Tasmania
Browse
vamplew-thesis.pdf (1.83 MB)

Recognition of sign language using neural networks

Download (1.83 MB)
thesis
posted on 2023-05-26, 07:46 authored by Vamplew, P
This thesis details the development of a computer system (labelled the SLARTI system) capable of recognising a subset of signs from Auslan (the sign language of the Australian Deaf community), based on the pattern classification paradigm of artificial neural networks. The research discussed in this work has two main streams. The first is the creation of a practical sign classification system, suitable for use within a sign language training system or other applications based on hand gestures. The second is an exploration of the suitability of neural networks for the creation of a real-time classification system with the ability to process temporal patterns. Sign languages such as Auslan are the primary form of communication between members of the Deaf community. However these languages are not widely known outside of these communities, and hence a communications barrier can exist between Deaf and hearing people. The techniques for recognising signs developed in this research allow the creation of systems which can help to eliminate this barrier, either by providing computer tools to assist in the learning of sign language, or potentially the creation of portable sign-language-to-speech translation systems. Artificial neural networks have proved to be an extremely useful approach to pattern classification tasks, but much of the research in this field has concentrated on relatively simple problems. Attempting to apply these networks to a complex real-world problem such as sign language recognition exposed a range of issues affecting this classification technique. The development of the SLARTI system inspired the creation of several new techniques related to neural networks, which have general applicability beyond this particular application. This thesis includes discussion of techniques related to issues such as input encoding, improving network generalisation, training recurrent networks and developing modular, extensible neural systems.

History

Publication status

  • Unpublished

Rights statement

The University is continuing to endeavour to trace the copyright owner(s) and in the meantime this item has been reproduced here in good faith. We would be pleased to hear from the copyright owner(s).

Repository Status

  • Open

Usage metrics

    Thesis collection

    Categories

    No categories selected

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC