The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
the ISPRS Congress in Amsterdam) presented detection of 3.1 Pulse-coupled neural networks
roads on aerial images with the use of Hopfield network for the
optimisation of the matching algorithm (Hu et al., 2000). A pulse-coupled neural network (PCNN) is a model of a
biological network, and specifically, a model of fragment of
However, the most popular application is the use of neural cat ' s sight network (Eckhom et al., 1990). Those networks are
networks for the classification of multi-spectrum image use< 3 to process digital images. By using them, it is possible to
contents. In one of the examples (Vieira, 2000), the author perform such operations as image segmentation, or contour
utilized those networks as one of classification methods. The detection. The theory and possibilities of application of those
ANN (Artificial Neural Network) method was used in addition networks were most fully described in (Lindblad et al., 2005).
to LCA, PTS, and MDR methods. The networks were also used
for the classification of specific INSAR-type images, where A pulse-coupled neural network is a single-layer network. It is
regions damaged as a result of earthquakes were detected (Ito et composed of neurons, and each of them is linked to one pixel of
al., 2000). Moreover, the networks were used to create thematic the input image. Fig. 1 shows the block diagram of a single
maps based on satellite data, with the use of the classification neuron. This is a simplified model of a PCNN network,
method (Barsi, 2000), in relation to data provided by Landsat. introduced by Kinser (Kinser, 1996) and named ICM
Other examples of utilization of neural networks for the (Intersecting Cortical Model). The simplification consists in the
classification of remote sensing images can be found in the elimination of one of feedback loops. The pulse-coupled neuron
work by Kamiya (Kamiya, 2000). is described by means of the following equations:
Another example of the use of artificial neural networks
there is no need to discuss it in this paper. Also computer
programs were prepared to facilitate the use of neural networks
even by people of a lesser experience (Duin, 2000; Matlab, A single neuron generates a pulse in the moment, when total
concerned automatic name placement on digital maps (Hong,
2000), or rectification of satellite images (Sunar et al., 2000).
(1)
There were also attempts at using neural networks for image
correlating, although the issue has so far been tested mainly in
the field of robotics. Some researchers (Pajares et al., 1998)
tried to solve the problem through the utilization of the network
teaching strategy, which was based on Self-Organizing Feature
Mapping (SOFM), being one of the Kohonen network variants.
kl
(2)
when
(3)
0otherwise
3. OUTLINE OF THE NEURAL NETWORK
TECHNIQUE
where:
The majority of neural network applications is related to a
broadly-understood classification. The methodology, which is
applied as a standard one, consists in preparing suitable
representations of image fragments and using them for the
classification of various network types. Most often that
classification employs a backpropagation-type network, which
is taught by means of a method "involving a teacher", or a
Kohonen's SOM network (Kohonen, 1997), which is an
example of teaching without supervision. In both instances, the
teaching of a network consists in changing the values of neuron
connection weights. Those changes take place as a result of
feeding successive representations, which have been collected
in the teaching set, to the network input, and applying proper
rules. In the case of backpropagation networks, the weight
changes depend on differences between response generated by
the network and the anticipated response, as defined by the
experimenter. In the Kohonen layer, after the feeding of a
teaching vector, the neuron, which has generated the highest
value, is determined. The weights of that neuron, as well as of
its neighbours are changed so that to approximate the value of
the input vector.
Sjj - excitation (input image pixel rescaled to the [0, 1]
range),
Fy - component involving the feedback
&ij - neuron threshold,
Yij - external state of the neuron (1 - pulse, 0 - no
pulse),
Wj - ij neuron neighbourhood coefficients, with kl
j k i coordinates for F
V F - gain amplification coefficient for F
n - iteration number («=!,...,N).
The theory, which describes the operation of those networks,
was presented in numerous publications (Arbib, 1995; Bishop,
1995; Haykin, 1999; Tadeusiewicz et al., 2007) so it seems that
Figure 1. Single neuron of the ICM pulse-coupled network
2006).
level of excitation of F exceeds the threshold value. After the
pulse has been generated, the threshold value rapidly increases,
and then gradually drops down to its rest value. In addition to S
image pixel value, also feedback signal from adjacent neurons
affects the F neuron excitation. During the subsequent
332