Abstract
The problem of unsupervised texture segmentation was studied and a texture segmentation algorithm was developed making use of the minimum number of prior assumptions. In particular, no prior information about the type of textures, the number of textures and the appropriate scale of analysis for each texture was required. The texture image was analysed by the multiresolution Gabor expansion. The Gabor expansion generates a large number of features for each image and the most suitable feature space for segmentation needs to be determined automatically. The two-point correlation function was used to test the separability of the distributions in each feature space. A measure was developed to evaluate evidence of multiple clusters from the two-point correlation function, making it possible to determine the most suitable feature space for clustering. Thus, at a given resolution level, the most appropriate feature space was selected and used to segment the image. Due to inherent ambiguities and limitations of the two-point correlation function, this feature space exploration and segmentation was performed several times at the same resolution level until no further evidence of multiple clusters was found, at which point, the process was repeated at the next finer resolution level. In this way, the image was progressively segmented, proceeding from coarse to fine Gabor resolution levels without any knowledge of the actual number of textures present. In order to refine the region-labelled image obtained at the end of the segmentation process, two postprocessing pixel-level algorithms were developed and implemented. The first was the mixed pixel classification algorithm which is based on the analysis of the effect of the averaging window at the boundary between two regions and re-assigns the pixel labels to improve the boundary localisation. Multiresolution probabilistic relaxation is the second postprocessing algorithm which we developed. This algorithm incorporates contextual evidence to relabel pixels close to the boundary in order to smooth it and improve its localisation. The results obtained were quantified by known error measures, as well as by new error measures which we developed. The quantified results were compared to similar results by other authors and show that our unsupervised algorithm performs as well as other methods which assume prior information.