Computational Intelligence - May 2016 - 17

I. Introduction

©GRAPHIC STOCK

owadays, the visual information handled by multimedia systems is growing significantly, in
part because of the technological advances. These have caused image capture devices to
be increasingly powerful and accessible in today's society, becoming essential elements of
everyday life. Another significant technological advancement has occurred in the field of
ccommunications,
ommunications, allowing these devices to share pictures and videos. All of this implies an increasing necessity of techniques for automatic image analysis and processing, enabling to efficiently
manipulate such information.
Images need to be represented computationally in order to be processed and analyzed by computers. From the computational point of view, an image is a dataset represented by a matrix of pixels. Each pixel represents color information, considered as a fundamental and representative
characteristic of visual content, and represented by a vector system. However, it is known that
humans use color terms when describing it [1]-[5], although there is no direct correspondence
between the color representation on a computer and the terms that humans use to identify them
(problem known as "semantic gap" [6]).
In this context, color modeling is a complex problem.
Among other reasons, colors are in general imprecise since it
is not possible in general to draw a clear boundary delimiting
colors (for example, the concept "red" is a gradual concept in
the color set, where the boundary between what is red and
not red is fuzzy). They are also subjective, because not everyone distinguishes or names in the same way the colors, and
context-dependent, because the same color can have different
meanings in different areas, even for the same person.
For instance, Figure 1, composed of a linear gradient
between two colors, shows the imprecise nature of color,
because it is not immediate to establish a precise boundary
separating both colors, but the boundary is fuzzy. Further,
there are colors in the gradient, especially in the central area,
for which a human could not tell which of the two color
names (associated with extreme colors) is the right one (possibly more than one, to a certain degree in each case). Indeed,
if those colors have to be named, the answer may be different
depending on the subject and context (e.g., a winemaker
could name them as "vermilion" and "purple", while a
greengrocer could use terms as "strawberry" and "aubergine"). In addition, in order to illustrate the subjective nature
of color, Figure 2 shows representative values assigned by different users to the color "red" in terms of points in RGB
space, where it can be noted that not every user assigns the
same representative value to the same color.
The examples above show that it is necessary to provide
mechanisms to model the color considering the imprecision,
subjectivity and context dependency. To do this, the most studied and used models (mainly to address the problem of imprecision) are fuzzy models based on the Theory of Fuzzy Sets
proposed by L.A. Zadeh [7]. Several authors agree on the fuzzy
nature of color, starting with the works of Rosch [8], [9], Kay
and McDaniels [10], where linguistic labels, represented by
fuzzy subsets of crisp colors, are used. Each fuzzy set comprises
a portion of the color space with fuzzy boundaries, defining a
fuzzy partition color space, so that each fuzzy subset of colors
corresponds to one of the color terms we use.
However, existing proposals in the literature present some
problems which, in our opinion, are not well solved. Current

may 2016 | IEEE ComputatIonal IntEllIgEnCE magazInE

Table of Contents for the Digital Edition of Computational Intelligence - May 2016