Page 205 -

Section 6.2 Pooled Texture Representations by Discovering Textons 173

dictionary 1

a Cluster 3
Learning 2 Dictionary

region

a Replace 1 3
Representing closest 2 3 Histogram
with
cluster
center

FIGURE 6.8: There are two steps to building a pooled texture representation for a texture
in an image domain. First, one builds a dictionary representing the range of possible pat-
tern elements, using a large number of texture patches. This is usually done in advance, us-
ing a training data set of some form. Second, one takes the patches inside the domain, vec-
tor quantizes them by identifying the number of the closest cluster center, then computes
a histogram of the different cluster center numbers that occur within a region. This his-
togram might appear to contain no spatial information, but this is a misperception. Some
frequent elements in the histogram are likely to be textons, but others describe common
ways in which textons lie close to one another; this is a rough spatial cue. This figure shows
elements of a database collected by C. Liu, L. Sharan, E. Adelson, and R. Rosenholtz, and
published at http: // people. csail. mit. edu/ lavanya/ research_ sharan. html .Fig-
ure by kind permission of the collectors.

Build a dictionary:
Collect many training example textures
Construct the vectors x for relevant pixels; these could be
a reshaping of a patch around the pixel, a vector of filter outputs
computed at the pixel, or the representation of Section 6.1.
Obtain k cluster centers c for these examples

Represent an image domain:
For each relevant pixel i in the image
Compute the vector representation x i of that pixel
Obtain j, the index of the cluster center c j closest to that pixel
Insert j into a histogram for that domain

Algorithm 6.2: Texture Representation Using Vector Quantization.