14 2. BASICS OF FEATURE DESIGN
density estimator in a regular grid [Felsberg et al., 2006]. e mathematical proof has basically
already been given in the context of averaged shifted histograms by Scott [1985].
In the work mentioned in the previous paragraph, channel coding is applied to the feature
domain, i.e., as a histogram approach. Obviously, it can also be applied to the spatial domain, i.e.,
as a signature approach. Combining both results in CCFM [Jonsson and Felsberg, 2009]. Both
SIFT and HOG descriptors can be considered as a particular variants of CCFMs, as will be
shown in Chapter 4. e CCFM framework allows generalizing to color images [Felsberg and
Hedborg, 2007a] and its mathematical basis in frame theory enables a decoding methodology,
which also includes visual reconstruction [Felsberg, 2010]; see Chapter 5.
Channel representations have originally been proposed based on a number of properties
(non-negativity, locality, smoothness; Granlund [2000b]). ese properties together with the
invariance requirement for the L
2
-norm of regularly placed channels [Nordberg et al., 1994]
imply the frame properties of channel representation [Forssén, 2004] and the uniqueness of
the cos
2
-basis function [Felsberg et al., 2015]. Irregular placement of channels, as suggested
by Granlund [2000b], obviously does not result in such a stringent mathematical framework,
but is particularly powerful for image analysis using machine learning and the representation of
less structured spaces, such as color. In that sense, color names [Van De Weijer et al., 2009] can
be understood as a non-regular spaced channel representation.
Similar to RGB color-space and intensity space, the non-negativity constraint for channel
representations implies a non-Euclidean geometry. More concretely, the resulting coefficient
vector lies in a multi-dimensional cone and transformations on those vectors are restricted to be
hyperbolic [Lenz et al., 2007]. is also coincides with observations made by Koenderink and
van Doorn [2002] that Euclidean transformations of image space (spatio-intensity space) are
inappropriate for image analysis.
A further conclusion is that the L
2
-distance is inappropriate to measuring distances in
these non-negative spaces. Still, many applications within image analysis and machine learning
are based on the L
2
-distance. More suitable alternatives, based on probabilistic modeling, are
discussed in Chapter 6.
2.5 LINKS TO BIOLOGICALLY INSPIRED MODELS
As mentioned in the previous sections, channel representations originate mainly from technical
requirements and principles, but were also inspired by biology [Granlund, 1999]. ey share
many similarities with population codes in computational neuroscience [Lüdtke et al., 2002,
Pouget et al., 2003], which are conversely mainly motivated by observations in biological sys-
tems.
To complete confusion, the concept of population codes developed historically from ap-
proaches that used the term “channel codes” [Snippe and Koenderink, 1992]. Channel (or pop-
ulation) codes have been suggested repeatedly as a computational model for observations in
human perception and cognition.