Classical Neural Networks - Neural Networks and Micromechanics - page 17

segmentation techniques is useful in assisting medical experts in the diagnosis and
treatment of tumors. The segmentation of single-channel magnetic resonance
images is a daunting task due to the relatively small amount of information avail-
able at each pixel site. This method has been validated on both simulated and real
images of volunteers and brain tumor patients. This is the first step in developing
a fully automatic segmentation method.
The other area where Kohonen maps found an application is for a distributed
measurement system for water quality monitoring [ 59 ]. Water quality monitoring
of rivers and seas represents an important task of life-quality assessment. This
monitoring is characterized by multi-parameter measurement capabilities. The main
parameters associated with water quality inspection can be classified in three cate-
gories: physical, chemical, and biological parameters. They use an advanced proces-
sing of data sensors based on auto-associative neural networks (Kohonen maps) in
order to offer a global water quality representation for a large monitored area.
One more interesting application of Kohonen maps is feature selection for
object extraction [ 60 ]. Selecting a set of features that are optimal for a given task
is a problem that plays an important role in pattern recognition, image understand-
ing, and machine learning. Li Pan et al. used Kohonen maps for continuous data
discretization in texture-recognition tasks. As the test task, they used tree recogni-
tion from aerial images.
Also, Kohonen maps were used for color image compression [ 61 ]. With the
development of multimedia technology and the Internet, image communication
including transmission, display, and storage at high speed has become increasingly
important. In this case, the hardware design for a neural-network-based color image
compression was developed. Compression using neural networks is advantageous
due to their features such as inherent parallelism, regular topology, and their
relatively small number of well-defined arithmetic operations involved in their
learning algorithms. So, VLSI implementation of Kohonen's map neural network
is well suited to color image compression due to its topological clustering property.
In this case, similar colors are clustered together and can be represented by one
color. The time complexity of the proposed scheme is linear in the image size. With
ASIC implementation the compression time is only a few milliseconds for images
of sizes up to 512

512 pixels.
Extended Kohonen networks were used for the pose control of microrobots in a
nanohandling station [ 62 ]. These are only a few examples among many others.

2.6.4 Cognitron and Neocognitron

Kunihico Fukushima developed and proposed the cognitron neural network [ 30 ],
which is an example of a hierarchical network [ 43 ]. It was initially proposed as a
neural network model of the visual system that has a hierarchical multilayered
architecture similar to the classical hypothesis of Hubel and Wiesel [ 63 , 64 ]. The
model consists of S- and C-cells. S-cells resemble simple cells of the primary