Use this URL to cite or link to this record in EThOS:
Title: Unsupervised texture segmentation using multiresolution Markov random fields
Author: Li, Chang-Tsun
ISNI:       0000 0001 3609 2486
Awarding Body: University of Warwick
Current Institution: University of Warwick
Date of Award: 1998
Availability of Full Text:
Access from EThOS:
Access from Institution:
In this thesis, a multiresolution Markov Random Field (MMRF) model for segmenting textured images without supervision is proposed. Stochastic relaxation labelling is adopted to assign the class label with highest probability to the block (site) being visited. Class information is propagated from low spatial resolution to high spatial resolution, via appropriate modifications to the interaction energies defining the field, to minimise class-position uncertainty. The thesis contains novel ideas presented in Chapter 4 and 5, respectively. In Chapter 4, the Multiresolution Fourier Transform (MFT) is used to provide a set of spatially localised texture descriptors, which are based on a two-component model of texture, in which one component is a deformation, representing the structural or deterministic elements and the other is a stochastic one. Experiments show that the algorithm is efficient in alleviating class-position uncertainty via data propagation across resolutions. However, the blocking artifacts of the segmentation results show that it is preferable to combine both class and position information so as to achieve smoother and more accurate boundary estimation. In Chapter 5, based on the same MFT-MMRF framework, a boundary process is proposed to refine the segmentation result of the region process proposed in Chapter 4. At each resolution, all the image blocks on either sides of the preliminary boundary detected in the region process are treated as potential boundary-containing blocks (PBCB's). The orientation and the centroid of the boundary-segment contained in each PBCB are calculated. The sequence of PBCB's are then modelled as a MRF and the interaction energy between each pair of neighbouring blocks is defined as a function of the 'distance' D between the centroids of the two boundary segments. During the stochastic relaxation process boundary/non-boundary labels are assigned to the PBCB's. Once the algorithm converges, the centroids of the identified true boundary blocks are connected to form the refined boundary which is propagated down to the next resolution for further refinement.
Supervisor: Not available Sponsor: Government of the Republic of China on Taiwan
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available
Keywords: QA Mathematics