Image Processing Reference
In-Depth Information
divided into 4 sub-blocks and the feature vector is defined as a ratio of intensities of
those sub-blocks. In the circle approach, proposed in [25], the image is first reduced
in dimension by Gaussian pyramid decomposition and every block is divided into
four concentric circles. The feature vector is calculated as a mean of the image pixel
value in each circular region of every block.
Wang et al. [26] introduced the first four Hu moments as features. The image
is first reduced in dimension by Gaussian pyramid decomposition, and the Hu mo-
ments are computed from the overlapping blocks of the low-frequency image. The
use of Zernike moments of degree 5 as features was proposed by Ryu et al. [19].
6.4
CA for CMF Detection
Cellular automata can be described as a discrete system that contains a regular grid
of cells. Each cell can be in only one finite-state determined by the previous states
of a surrounding neighbourhood of cells (reference on CA). The use of cellular
automata for image processing is interesting because of the property that very simple
rules can result in very complex behaviour. Detection of CMF by CA is based on
the fact that similar areas in an image should produce similar rules.
The basic approach can be described as a variation of block-based methods for
CMFD with a new set of feature vectors. In this approach, the feature vector for
each block is obtained using a CA to generate a set of rules that describe the texture
of that block [23]. This process can be described as a selection of a subset of rules
from all possible rules.
Applying a CA on a greyscale image results in the combinatorial explosion in the
number of possible rules, because a whole range of image intensities (256 levels)
is used as cell states. For example, using a neighbourhood of 8 pixels for learning
rules, leads to 256
8
possible rules. Moreover, the large number of possible rules
leads to an even larger number of possible subsets of those rules, which makes it
difficult to learn and represent rules efficiently. A reduced description of a neigh-
bourhood can be accomplished by a proper binary representation of image. In that
case only two values (0 and 1) are used as cells states leading to a compact descrip-
tion. For comparison, a neighbourhood of 8 pixels on binary image generates only
2
8
possible rules. It is clear that a binary representation of image is more suitable
for this purpose, but the problem that arises is to find a proper binary representation.
In the following section, we describe several possible solutions.
Another important part of CMFD with CA is the proper selection of neighbour-
hood. Some tasks can be solved by using simple 1D neighbourhood, but in cases of
some transformations, complicated versions of neighbourhood have to be applied.
6.4.1
Representation of Image in Binary
Applying a CA to a grayscale image requires taking the whole range of intensities as
a cell states which leads to a large number of possible rules, and a large number of
possible subsets of rules. One way to avoid that is to represent the grayscale image
Search WWH ::
Custom Search