Mapping Associations - Mapping Scientific Frontiers: The Quest for Knowledge Visualization - page 97

high-dimensional space. To be precise, the dimensionality of the space is the number
of pixels on an image: 64 64 D 4,096. The MDS and PCA techniques introduced
later in this chapter can be applied to such sets of images.
The key issue in content-based image retrieval (CBIR) is how to match two
images according to computationally extracted features. Typically, the content of
an image can be characterized by a variety of visual properties known as features.
It is common to compare images by color, texture, and shape, although these entail
different levels of computational complexity. Color histograms are much easier to
compute than a shape-oriented feature extraction.
Computational approaches, on the other hand, typically rely on feature-extraction
and pattern-recognition algorithms to match two images. Feature-extraction algo-
rithms commonly match images according to the following attributes, also known
as query classes:

• Color
•Texture
•

Shape

• Spatial constraints

Swain and Ballard ( 1991 ) matched images based solely on their color. The
distribution of color was represented by color histograms, and formed the images'
feature vectors. The similarity between a pair of images was then calculated using
a similarity measure between their histograms called the normalized histogram
intersection . This approach became very popular due to its robustness, computa-
tional simplicity, and low storage requirements. A common extension to color-based
feature extraction is to add textural information. There are many texture analysis
methods available, and these can be applied either to perform segmentation of the
image, or to extract texture properties from segmented regions or the whole image.
In a similar vein to color-based feature extraction, He and Wang ( 1990 )useda
histogram of texture, called the texture spectrum . Other types of features include
layout and shape.
In the following example, we visualized a set of 279 visualization images.
The majority of these images are synthetic graphics generated by computer or
screenshots of information visualization systems. The size, resolution, and color
depth of these images vary. Images were grouped together by a human user in order
to provide a point of reference for the subsequent automatically generated models.
We asked the user to group these images according to their overall visual similarity,
but no specific guidelines were given on how such similarity should be judged.
Similarity measures between these images were computed by the QBIC system
(Flickner et al. 1995 ). The three networks correspond to similarities by color, layout,
and texture. We expected that images with similar structures and appearances should
be grouped together in Pathfinder networks.
Figure 3.6 is the screenshot of the visualization. The Pathfinder network was
derived from similarities determined by color histograms. The layout of the
visualization is visually appealing. Several clusters of images have homogenous
colors. The largest image cluster includes images typically with line-drawing-like