![](bg2.png)
2 MATERIALS AND METHODS
2.1 Theoretical Background
Clustering - is the automatic partitioning of a set of
elements into groups according to their similarity.
Elements of the set can be anything, for example,
data or characteristics vectors. Themselves groups
are also called clusters (Tryon, 1939).
In our case, using algorithms of cluster analysis
will be the identification of ore minerals by color
and texture characteristics of color-coded minerals
identified in images taken in reflected light using a
microscope (Panteleev, Egorova and Klykova,
2005).
The proposed method consists in measuring the
intensity of the three spectral components of the
reflected light (red, green, blue) in each pixel of the
investigated surface (frame) of a rock sample. Rock
sample is subjected to a pre-treatment and
preparation of the surface to be scanned, for example
in the form of ore.
In most cases preparation of the samples
represented by the following general form:
- Cutting of the sample;
- Grinding;
- Polishing.
After the sample has been prepared, it is
analysed with a microscope. A laboratory
microscopic picture reflects the structural features of
objects (color, texture, space and so on), determines
the results of mineralogical analysis.
Theoretically possibility to determine the mineral
ore targets on the microscopic image substantiated
by author M.P. Isayenko ((Isayenko, Borishanskaya
and Afanasyev, 1986).
In this paper it describes the algorithm for
automatic segmentation of color images of rocks,
using the methods of cluster analysis. There are
results of studies different color spaces for clustering
k-means (Huang, 1998).
In general, the K-means method segments the
image on K different clusters (areas) located far
away from each other based on certain criteria
(Odell and Duran, 1974).
As such a characteristic can be selected color
(the values of all three components simultaneously
RGB) color and geometric distance at the same time,
etc. By default, the implementation of this method is
applied to states Euclidean metric (Mandel, 1988).
Segmentation method "K- means" is
implemented through a two-step algorithm that
minimizes the sum of distances "point-to- centroid"
obtained by summing over all K clusters. Another
words, the purpose of the algorithm is to minimize
variability within clusters and maximize variability
between clusters (Ryzin, 1977).
Algorithm starts with a randomly selected cluster
centroid position, and then changes the ownership of
points (objects) to clusters, i.e. point moves from
one cluster to another in order to get the most
significant result.
During the first phase on each iteration all points
are rearranged so that they are positioned as close as
possible to their centroids, and then converted
coordinates centroids of each cluster. This part of the
algorithm allows to find quickly, but only an
approximately a solution to the problem of
segmentation, which is the starting point for the
second phase.
During the second stage of the algorithm points
are individually subjected to rearrangement in case it
reduces the sum of the distances, and the coordinates
of the centroids clusters after rearrangement
recalculated for each point. Each iteration during the
second stage consists of only a single pass through
all the points.
After completion of the segmentation algorithm
described program may provide additional
information such as:
- Sum of distances "point-to-centroid";
- Coordinates of centroid as well as some other
data.
Algorithm K-method can converge to a local
optimum, when the separation points move any
point to another cluster increases the resultant sum
of the distances. This problem can be solved only by
a reasonable (successful) choice of initial points
(Odell and Duran, 1974).
2.2 Color Image Segmentation
Algorithms
Segmentation is the process of dividing an image
into regions. Color segmentation in the vector space
RGB is as follows. Suppose that our goal is to
allocate objects in the image RGB, the color of
which lies within a certain range. Having some
representative sample vectors, we are interested in
having the color; we obtain an estimate of the
"average" of color you want to select. Let this
average color RGB denotes a column vector of T.
The problem of segmentation is to classify each
pixel RGB image and determine it belongs selected
"average" color-class or not. It is necessary to have
some measure of similarity of colors to implement
such a comparison (Martin, Fowlkes and Malik,
2004).
SIGMAP2014-InternationalConferenceonSignalProcessingandMultimediaApplications
166