BibTeX entry

@PHDTHESIS{200305Derek_Charles,
  AUTHOR={Derek Charles},
  TITLE={Algorithmic and learning based filtering techniques with
    application to colour image noise suppression and enhancement},
  SCHOOL={University of London},
  MONTH=May,
  YEAR=2003,
  URL={http://www.bmva.org/theses/2003/2003-charles.pdf},
}

Abstract

This thesis investigates the principal local-window noise-removal techniques in the spatial domain using linear and non-linear filters for a variety of colour image representations and general and specialized image classes. The thesis starts by describing the nature of colour, general image processing capture and storage techniques, and the manner in which image contamination can occur. A software framework was designed and implemented in which the image filtering algorithms were developed: its architecture and capabilities are outlined. After a review of the current state-of-the-art in the field of vector filtering, performance comparisons of well known nonlinear and hybrid filters are achieved using established noise metrics; the validity and efficacy of these noise metrics are also debated. A number of new filters based on the vector median filter are introduced with a discussion of their relative merits. The new distance-weighted median filter is shown to be superior in its noise suppression performance characteristics using the NMSE criterion. Particular regard is given to an extension of the truncated-median filter of Davies (1988), so that it can cope with colour images, with a discussion on its performance with highly contaminated images and a comparison of the distortions introduced by the new filter and those of the vector median filter. The new version of the truncated median filter is shown to be quite remarkable in its capability for extraction of partially hidden patterns in colour images. The performance of the vector median filter and the variants described are contrasted with the capabilities of artificial neural networks (ANNs), in particular the multilayer perceptron, and the applicability of ANNs to colour image filtering is compared with previously published work on their use in greylevel images. The thesis contains a considerable amount of review of previous work in the subject area, together with a substantial list of references.