Statistical Image Modeling and Processing Using Wavelet-Domain Hidden Markov Models - Nonlinear Signal and Image Processing: Theory, Methods and Applications - page 370

FIGURE 10.15
The sample images (64 × 64, 8 bpp) of the 55 Brodatz textures used in the experiment.

likelihood functions f (IMM, HMT,

and HMT-3S) describe the data w . For simplicity, we consider the case in which
the prior probabilities to the texture classes are equal, and where the goal is
to minimize the overall probability of classification error. The optimal deci-
sion becomes the maximum likelihood rule, which is to choose the class that
makes the observed data most likely, i.e.,

( w | θ) that measure how well the models θ

c

C ML = arg max f ( w | θ ). (10.39)

c ∈{ 1 , ... ,N c }

The simulation of texture classification in this work is based on 55 Brodatz
textures of a size of 640

×

640 (Figure 10.15).* There were two reasons for us
to choose these 55 textures for this work. One was that the overall percentage
of correct classification (PCC) of the WES alone is less than 80% on this set
of textures, thus creating a difficult problem in examining the IMM, HMT,
and HMT-3S in terms of the accuracy of texture characterization. The other
was that the classification experiment here is conducted on 64 64 texture

samples. We excluded the textures that are nonhomogeneous in this small
size, thus providing a reasonable test environment.
Prior to the classification experiment, four sets of texture features or models,
i.e., WES, IMM, HMT, and HMT-3S, are obtained and stored for each texture.

×

* These textures are obtained from the Brodatz database at http://www.ux.his.no/ ∼ tranden/

brodatz/.