|M.Sc Thesis||Department of Electrical Engineering|
This study deals with the application of multi-resolution statistical models to the tasks of texture classification and segmentation. These tasks are at the core of many applications such as document processing and medical image processing. In this work
we choose to characterize textures using the statistical distribution of their multi-resolution
representation’s coefficients, and employ various statistical models for this
purpose. First we describe the statistical models which we employ in this work, and
make the distinction between models that assume that every coefficient of the
multi-resolution representation is statistically independent as well as models that can
account for statistical dependencies between the coefficients.
In this work we propose new feature statistics for wavelet based texture classification,
and an appropriate feature weighting method. We show that the classification results
of the new method are superior to those reported recently in the literature using the same
experimental tools and testing procedures.
When considering segmentation of multi-textured images, it is common to enforce
additional prior information for the characterization of the texture classes. Two
such prior models which are employed in this work are the level-set framework that
can enforce smooth texture boundaries, and the Markov random fields framework
that can enforce the smoothness of the different texture regions. Using the level-set
framework we propose a new multi-scale level-set supervised texture segmentation
scheme, which employs a coarse to fine strategy, thus we show that the new scheme
may have advantages over presently available level-set supervised texture
segmentation schemes, where the segmentation is performed on a single level.
Finally, we employ the Markov random fields framework and consider the use of a
pre-processed feature space common in the literature for unsupervised texture
segmentation using Gaussian Markov random fields. We show that the pre-processed
feature space has certain advantages for this task due to its Gaussian statistics.