Multilayer perceptrons for image data compression and speech recognition

Huang, Shih-Chi. 
 University of Notre Dame ProQuest Dissertations Publishing,  1991. 9210348.

Abstract (summary)

This dissertation addresses fundamental issues of multilayer perceptrons (MLP) and their applications in two technical areas, namely, image data compression and speech recognition. Image data compression plays an irreplaceably important role in many real-world communications and signal processing applications. Typical such examples include high-definition television, video-phone systems, teleconferencing systems, fax machines, and many others. On the other hand, the development of advanced large-vocabulary speaker-independent continuous speech recognition technology can be used as the phonetic keyboard of computers and will make many sophisticated systems much more user-friendly. Solutions to the problems of image data compression and speech recognition mandate a system which can accomplish a great deal of computation within a short-time interval and has a high degree of flexibility to deal with numerous possible inputs. Artificial neural networks provide answers to both issues. This dissertation focuses its attention to one kind of artificial neural networks, i.e., multilayer perceptrons (MLP). The adaptability of an MLP makes it capable of modeling many systems, makes it inclusive of other kinds of artificial neural networks and raises many issues to be investigated. The first half of this dissertation will be concerned with some fundamental issues for multilayer perceptrons, including capability of a fixed-size MLP and learning algorithms to determine the connection weights for an MLP. Problem formulations and overviews of literature survey to highlight the significance of these investigations will be given. The second half of this dissertation will then investigate image data compression and speech recognition. This dissertation proposes a vector quantization technique employing the entropy concept for image data compression and an extended subspace method for speech recognition. Both of these methods will be shown to overcome many implementational difficulties encountered by conventional methods. It is further proposed to implement the resulting algorithms using MLPs. Speed and flexibility provided by MLPs make the proposed algorithms more appealing than conventional approaches to many modern applications.

Indexing (details)

Electrical engineering;
Computer science;
Artificial intelligence
0544: Electrical engineering
0984: Computer science
0800: Artificial intelligence
Identifier / keyword
Applied sciences; data compression
Multilayer perceptrons for image data compression and speech recognition
Huang, Shih-Chi
Number of pages
Degree date
School code
DAI-B 52/11, Dissertation Abstracts International
Place of publication
Ann Arbor
Country of publication
United States
Huang, Yih-Fang
University of Notre Dame
University location
United States -- Indiana
Source type
Dissertation or Thesis
Document type
Dissertation/thesis number
ProQuest document ID
Database copyright ProQuest LLC; ProQuest does not claim copyright in the individual underlying works.
Document URL