 
Summary: ENEE 739J Assignment 2
Submitted by Amit Agrawal
1 Problem Statement
Unsupervised and Supervised Texture Segmentation using Markov Random
Fields
2 Supervised Texture Segmentation
In supervised texture segmentation, the classes of textures are known before
hand. In other words, model parameter estimation is easier. The texture is
assumed to be modelled by a 5th order GMRF. From the given image,regions
of each texture are extracted. From the interior pixels for each such regions, a
5th order GMRF is fitted using least squares.
Once parameter estimation is done, the energy function U corresponding to
each pixel for every texture is estimated. For each pixel a k*k window is taken
and the energy function is calculated as given in [1]. Using this prior, initial
labels for each pixel can be calculated as the one which minimizes U among all
texture classes for that pixel.
The labels or texture classes are also modelled as 5th order GMRF. The
energy function UL for the labels is assumed to be the form given in 1 with beta
= 1.4.
Bias w(L) for the textures plays a very important role. It is difficult to
