7. Istanbul 2004
ı many pattern
sing
X COOCCUrrence
f possible pixel
on will be easily
2). Hence, it is
y values into à
scalar or vector
ature extraction
ties which have
ransformed into
tion (Gonzales,
as a sample of
Is of two pixels
sian coordinates
be chosen as à
ce matrix N is
FAX, y+Ay)=j}
ach band have
his generalized
l cooccurrence
the generalized
1 the gray level
: of the second-
e measurement
urrence matrix
age are highly
tered along the
(low have been
ry spread about
(9)
(10)
(11)
(12)
(13)
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
LA LA
Sy = la —b| P(a,b) (14)
a=0h=0
Inverse difference:
SS P(a,b) (15)
a=0h=0 1 + (a = bY
Energy:
LA LA
S, 7S Vip(a 5E (16)
a=0b=0
Entropy:
ue Plato Plas) UD
0
a=0bh=
The measures mentioned above to specify the texture have
been applied in this study. The utilization of the second-order
measures for texture analysis is considered in section 5.
3. SEGMENTATION
Segmentation of an image entails the division or separation of
an image into regions of similar attribute. The most basic
attribute for segmentation is image luminance amplitude for a
monochrome image and color components for a color image.
Image edges and texture are also useful attribute for
segmentation.
Artificial neural networks are being increasingly employed in
various fields such as signal processing, pattern recognition,
medicine and so on. An approach to compress color image
+ data using a SOM is proposed, with the intention to generate
a code table for color quantization for data transformation
(Godfery, 1992).
The present work deals with the color subset and describes an
attempt to segment color image based on a self organizing
neural network, which is able to recognize main components
(chromaticities) present in the image. The next step is about
identifying these components and classifying the image data
based on the detected classes. The next section describes the
self-organizing maps fundamental.
3.1 Self-organizing learning
Self-organized learning or unsupervised learning paradigms
are forms of the cluster analysis. Clustering generally
describes a regression solution that attempts to optimally
partition an input space of dataset of N elements into a
compact representative set of K cluster centers, where K« «N.
For example the input space data may represent the classified
pixels in an image, and the cluster represent color image
segments.
3.1.1. Self-organizing maps
The self-organizing neural network, also known as Kohonen
network (Kohonen, 1989), is a network that incorporates a
topology scheme, i.e., it takes into account the topological
Structure among neurons. The input signals are n-tuples.
There is a set of m cluster units. Each input unit is fully
connected to all output units which response differently to the
input pattern. At the time of each input at the training phase,
the cluster unit with weights that best match the input pattern
is selected as the winner (usually in a minimum Euclidean
distance sense). This winning unit and the neighborhood
around it are then updated in such a way that their internal
119
weights be closer to the presented input. The adopted
updating factor is not equal for all neurons, but stronger near
the winning unit, and decreasing for more distant units.
Figure 1 shows the basic structure of self-organizing maps. It
shows input components (white circles) connected to all
cluster units (shaded circles). The units can assume any
spatial distribution, which are usually linear or planar arrays.
Weights are associated to each connection. With time, the
gain factor must be reduced together with the neighborhood
decrease in size.
Cluster units ©
Input layer
Figure 1. Self-organizing map
During the learning phase the node weights are changed in an
ordered manner in such a way that the main image features
tend to be organized according to topological distribution in
the network. Adjacent nodes response similarly, while distant
nodes respond diversely. The convergence of the feature in
the self-organizing map occurs considering some limitations
on the gain factor while updating the weights (Yin, 1995).
4. NEURAL NETWORK CLASSIFICATION
Classification is the process of sorting pixels into a finite
number of individual classes or categories of data based on
their original values. If a pixel satisfies a certain set of
criteria, then the pixel is assigned to the class that
corresponds to that criterion.
The most widely used neural. classifier is multilayer
perceptron network which has been extensively analyzed and
for which many learning algorithms have been developed.
The MLP belongs to the class of supervised neural networks.
The multi-layer perceptron neural network model consists of
a network of processing elements or node arrangement in the
layers. Typically it requires three or more layers of
processing nodes: an input layer which accepts the input
variables used in the classifier procedure, one or more hidden
layers, and an output layer with one node per class. Number
of neurons in the input layer depends on the features vector
and in the output layer is based on the number of classes. In
this study, a three-layer network has been constructed with 72
neurons in input layer, 40 neurons in hidden layer and 5
neurons in output layer.
There are several training algorithms for feed forward
networks. All of these algorithms use the gradient of the
performance function to determine how to adjust the weights
to optimize performance. The gradient is determined using a
technique called back propagation, which involves
performing computational backwards through the network.
In this paper, resilient back propagation training algorithm
has been used. It is a training algorithm that eliminates the
harmful effect of having a small slope at the extreme ends of
the sigmoid "squashing" transfer functions (Riedmiller,
1993). It is faster and more accurate than the standard back
propagation algorithm for training. It can converge from ten
to one hundred times faster than the standard algorithm using
