Original title: Segmentace obrazových dat
Translated title: Image Segmentation
Authors: Mikeš, Stanislav ; Haindl, Michal (advisor) ; Scarpa, Giuseppe (referee) ; Jan, Jiří (referee)
Document type: Doctoral theses
Year: 2010
Language: eng
Abstract: Image segmentation is a fundamental part in low level computer vision processing. It has an essential influence on the subsequent higher level visual scene interpretation for a wide range of applications. Unsupervised image segmentation is an ill-defined problem and thus cannot be optimally solved in general. Several novel unsupervised multispectral image segmentation methods based on the underlaying random field texture models (GMRF, 2D/3D CAR) were developed. These segmenters use efficient data representations that allow an analytical solutions and thus the segmentation algorithm is much faster in comparison to methods based on MCMC. All segmenters were extensively compared with the alternative state- of-the-art segmenters with very good results. The MW3AR segmenter scored as one of the best available. The cluster validation problem was solved by a modified EM algorithm. Two multiple resolution segmenters were designed as a combination of a set of single segmenters. To tackle a realistic variable lighting in images, the illumination invariant features were derived and the illumination invariant segmenter was developed. For the proper evaluation of segmentation results and ranking of algorithms, a unique web-based texture segmentation benchmark was proposed and implemented. It was used for comprehensive...
Keywords: benchmark; image segmentation; texture

Institution: Charles University Faculties (theses) (web)
Document availability information: Available in the Charles University Digital Repository.
Original record: http://hdl.handle.net/20.500.11956/26218

Permalink: http://www.nusl.cz/ntk/nusl-281305


The record appears in these collections:
Universities and colleges > Public universities > Charles University > Charles University Faculties (theses)
Academic theses (ETDs) > Doctoral theses