Includes bibliographical references. / In the first part of this dissertation, we present a detailed description of the eigenface technique first proposed by Sirovich and Kirby and subsequently developed by several groups, most notably the Media Lab at MIT. Other significant contributions have been made by Rockefeller University, whose ideas have culminated in a commercial system known as Faceit. For a different techniques (i.e. not eigenfaces) and a detailed comparison of some other techniques, the reader is referred to [5]. Although we followed ideas in the open literature (we believe there that there is a large body of advanced proprietary knowledge, which remains inaccessible), the implementation is our own. In addition, we believe that the method for updating the eigenfaces to deal with badly represented images presented in section 2. 7 is our own. The next stage in this section would be to develop an experimental system that can be extensively tested. At this point however, another, nonscientific difficulty arises, that of developing an adequately large data base. The basic problem is that one needs a training set representative of all faces to be encountered in future. Note that this does not mean that one can only deal with faces in the database, the whole idea is to be able to work with any facial image. However, a data base is only representative if it contains images similar to anything that can be encountered in future. For this reason a representative database may be very large and is not easy to build. In addition for testing purposes one needs multiple images of a large number of people, acquired over a period of time under different physical conditions representing the typical variations encountered in practice. Obviously this is a very slow process. Potentially the variation between the faces in the database can be large suggesting that the representation of all these different images in terms of eigenfaces may not be particularly efficient. One idea is to separate all the facial images into different, more or less homogeneous classes. Again this can only be done with access to a sufficiently large database, probably consisting of several thousand faces.
| Identifer | oai:union.ndltd.org:netd.ac.za/oai:union.ndltd.org:uct/oai:localhost:11427/14381 |
| Date | January 1998 |
| Creators | Muller, Neil Leonard |
| Publisher | University of Cape Town, Faculty of Science, Department of Mathematics and Applied Mathematics |
| Source Sets | South African National ETD Portal |
| Language | English |
| Detected Language | English |
| Type | Master Thesis, Masters, MSc |
| Format | application/pdf |
Page generated in 0.0026 seconds