Multiresolution MRF-Based Texture Segmentation Using the Wreath
Product Transform Phase
G. Mirchandani and X.Luo
ICASSP2001, Salt Lake City, Utah
(Paper (pdf) / Paper (ps))
In this paper, we bring local phase and multiresolution analysis
to the texture segmentation problem. A Markov random field
characterization is still employed, except it is used to model
{\em phase} correlations rather than {\em intensity} correlations.
Since statistical characteristics of phase are typically quite
different to those of intensity, there exists the potential for
creating greater discrimination in its feature space. We apply the
Wreath Product Transform and use the phase at higher scales to
initiate the segmentation process. For textures defined by
homogeneous regions of dominant local edges, we see that the new
algorithm yields better segmentation than that obtained through
conventional multiresolution algorithms based on lowpass data.