1

A REVIEW ON NEURAL NETWORK-BASED IMAGE

SEGMENTATION TECHNIQUES

CATALIN AMZA

De Montfort University
Mechanical and Manufacturing Engineering
The Gateway
Leicester, LE1 9BH
United Kingdom
[email protected]
Tel.: +(44) 07715171304
Fax: +(44) (0)(116) 2577099

Abstract
Image segmentation is the most important step in modern computer vision. Its output is
crucial for all the other stages of computer vision. The literature is very rich in segmentation
techniques and neural-network based methods have been applied successfully due to their
signal-to-noise independency, their ability to achieve real-time results and the ease of
implementing them with massive VLSI processors.

Keywords: image segmentation, neural networks.

Introduction
Any image can in general be described by a two-dimensional function f(x,y), where x
and y represent the spatial coordinates and f(x,y) the value at that location. Depending on the
type of image, the value f(x,y) can be either light intensity, temperature (for thermal images),
intensity of X-rays (for X-ray images), intensity of radio-waves (for nuclear magnetic
resonance images MRI), depth for range images, etc.
A digital image can be viewed as a two-dimensional matrix. Each element of the matrix is
called a pixel (x,y) and it represents a discrete value of the feature intensity value:
U
n y
m x
y x f I
, 1
, 1
) , (

,
(1)
where m and n are the dimensions of the image.
2
Most of the techniques discussed below are designed for light intensity images, therefore, in
the subsequent paragraphs we shall mainly refer to f(x,y) as the grey value of the pixel at
location (x,y).
Image segmentation is one of the most important processes in modern computer
vision. It involves partitioning the image into meaningful segments. It is the process by which
a computation translates the original image description i.e. an array of grey levels into
segments with uniform and homogenous characteristics. They should correspond to structural
units (objects) in the scene.
Thus, using (1), the segmentation can be described as partitioning the image I into S
1
, S
2
, ,
S
n
segments of objects so that
U
n
i
i
S I
1

(2)
and
I
j i
j i
S S

,
(3)
where n is the number of objects of interest.
There are many applications of the image segmentation. From the medical field to
robotics, image segmentation has played and plays an important role. For instance, for the
automated detection of cancerous cells from mammographic images, segmentation followed
by recognition or classification is required. Another example is that of automatic non-
destructive testing techniques, such as automatic inspection of welding, castings, detection of
foreign bodies within food products [1,2,3], etc. Such techniques involve the segmentation of
the image, and detection (recognition) of possible anomalies or foreign bodies within.
Therefore the output of such a system is, in most of the cases, directly dependant of the
segmented output of the original image.
Despite this, image segmentation remains a difficult problem due to lack of a general
mathematical model. A comprehensive number of techniques have been proposed in the
3
literature and are briefly described later in the paper. Selection of one method over another
solely depends on the type of image that needs to be segmented.
There is no universal method that can be successfully applied to all types of images.
Moreover, even the process of choosing the right segmentation method is a very difficult
process. Methods developed for one type of image might not be applied to different images,
or applied with low performances.
Segmentation techniques can be placed into three classes:
Classical algorithms mostly based on mathematical or statistical methods
Artificial Intelligence techniques
Other techniques which either crossover or fall into none of the first two categories.
The classical algorithms include characteristic histogram thresholding, edge/boundary
detection, region extraction or region growing, relaxation, semantic and syntactic approaches.
Segmentation methods are applied from the artificial intelligence field, especially using
neural networks approaches.
Previous image segmentation reviews [4], [5] and [6] only surveyed classical techniques;
neural network techniques have not been considered in these. A brief review of ANNs
methods used for segmentation is given in [7] and [8]. Pal and Pal [7] classified image
segmentation techniques as: a) grey level thresholding; b) iterative pixel classification; c)
surface based segmentation; d) segmentation of colour images e) edge detection. Pal and Pal
[7] reported neural networks based approaches used only for iterative pixel classification. [8]
and [9] reviewed only FFNN-based techniques used for segmentation of MRIs.
The present survey is intended to be a more comprehensive study of the existing
neural-network-based segmentation techniques. Due to the extensive number of segmentation
techniques reported in the literature, this survey is a selective one. Because most of the
4
methods in the literature can be applied or extended easily to colour images, in the following
discussion we will refer only to grey level images.

Artificial Neural Network-Based (ANN) techniques
Since 1990, artificial neural networks have come to be used as a different approach
for image segmentation. Their properties, such as graceful degradation in the presence of
noise, their ability to be used in real-time applications and the ease of implementing them
with VLSI processors, led to a booming of ANN-based methods for segmentation. Almost all
types of neural networks have been applied with a different degree of success. The mostly
used being Kohonen and Hopfield ANNs. In the present review, we will present methods
based on feed-forward back-propagation neural networks (BPNN or FFNN or MLFF or
MLP), Kohonen self organising maps (SOMs), Hopfield neural networks (HNN), constraint
satisfaction neural networks (CSNN), oscillatory networks (LEGION) and pulse-couple
neural networks (PCNN).
An artificial neural network is a simulation of a real nervous system. It consists in a
number of neurons that communicate with each other. This artificial computational model of
a real nervous system was proposed in 1943 by McCulloch and Pitts. This paper will not
attempt to describe ANNs in detail. More details about Neural Networks are presented in
[10], [11], [12] or in the referenced papers.
The NN-based image segmentation techniques reported in the literature can mainly be
divided into two categories: supervised and unsupervised methods (see figure 1).
Supervised methods require expert human input for segmentation. Usually this means
that human experts are carefully selecting the training data that is then used to segment the
images. Unsupervised methods or clustering processes are semi or fully automatic. User
intervention might be necessary at some point in the process to improve performance of the
5
methods, but the results should be more or less human independent. An unsupervised
segmentation method automatically partitions the images without operator intervention.
However, these architectures might be implemented using application specific a priori
knowledge at design time, i.e. anatomical, physical or biological knowledge.

Fig.1 NN-based image segmentation techniques

Supervised Techniques
Supervised segmentation techniques (figure 2) are based on human or operator
knowledge to select training images and manually segment them into k regions. Each region
is assign with a label and the proposed architecture is trained using the selected images as
training data. The method is then able to segment similar images. Labels are assigned to the
regions according to the knowledge stored in the NN architecture used.

Fig.2 Supervised NN-based image segmentation techniques
6

Multi-Layer Perceptron (MLP)
The simplest NN architecture used for image segmentation is the multi-layer
perceptron (MLP). Blanz and Gish [13] had successfully applied a three-layered perceptron
to segment grey-level images. Their BPNN approach transformed the image segmentation
problem into a pixel classification problem. The input vector for the ANN consisted of the
vector of features extracted from every pixel, and the output vector was the vector of classes
desired for segmentation. A standard back-propagation (BP) learning algorithm was chosen
to train the network.
Lawson and Parker [14] detailed a very similar approach to image segmentation of
industrial radiographic images using back-propagation algorithm. Waard [15] described a
MLP trained to classify pixels in a text segmentation application. Another back-propagation
neural network (BPNN) was implemented by Silverman and Noetzel [16] for detection of
tumours in ultrasound medical images. MR and CT images were also segmented using BPNN
[17,18,19].
A supervised method based on BPNN was used by Raff et al. [20] to segment the grey
matter, white matter and cerebrospinal fluid from spin-echo magnetic resonance images. The
histogram of the image was used to train a BPNN. The output of the architecture was the
percentage of grey or white matter in the image.
A different method was proposed by Babaguchi et al. [21]. Here, a binary image is
segmented by automatically selecting a threshold by a BPNN. The segmentation problem is
no longer transformed into a pixel classification problem, but into an optimal thresholding
selection problem. The inputs to the ANN are the histogram values and the output is the
appropriate value of the threshold.
7
A different learning algorithm was used by Hall et al. [22]. They have used fuzzy-
clustering techniques and a MLP trained with a cascade correlation-learning algorithm in
segmenting magnetic resonance brain images. This approach solves the problem of speed and
computational overhead that BPNN suffer from. The speed is increased by as much as two
times, but in the detriment of cost of acquiring training data.
Coppini et al. [23] described a system based on integrating a priori anatomical
knowledge and a feed-forward BPNN for the segmentation of target structures in
tomographic images and of lung nodules in standard projection radiography. Their approach
was inspired by the anatomical world: three major blocks were used to segment each type of
structure. The input was the retina. Then, the structures were localised by the Attention
Focuser (AF) and its findings were reported to the output of the system, called the Region
Finder (RF). All three blocks were implemented as BPNN, with different parameters. The
same approach is described by Valli et al. [24]. Even though their performance was around
95% correct pixel classification, the presence of noise in the input images was not
investigated at all. Their conclusion was that a priori anatomical knowledge could help
improve the neural-network based segmentation of medical images.
Wang et al. [25] investigated a NN used for segmentation of images containing
Gaussian noise and outliers. Their approach consisted of two stages: initial and refinement
segmentation. The image is represented by a FFNN so that neurons with the same state at one
time represent the same region (the same object/segment). An energy function is defined for
the network as being the negative of the between-class variance (see [26]):

1
0
2
) ( ) (
k
k k
p V E ,
(4)
8
where p
k
is the probability of the k
th
region,
k
are the mean grey-level for the k
th
region and
is the mean grey-level for the image. It represents the goodness of the segmentation and its
minimising will lead to the initial segmentation of the image into two subsets. The refining
segmentation procedure is achieved by a self-improvement neural network in that it learns
from itself, and improves itself by a retrieval procedure. Their architecture, called a multi-
layer logic neural network, was able to perform the learning algorithm in one pass. Even
though this method performs well in the presence of noise, it cannot be used when the
difference of grey-level within one segment is high i.e. texture regions.
Shiranita et al. [27] proposed a BPNN for implementation of a meat-quality grading
system. A three layered MLPNN was trained to classify pixels as fat-pixel pattern or
muscle-pixel pattern. A binarization was then performed on the image.

Kohonen Neural Network
Haring et al. [28] developed a Kohonen neural network [29] for multi-scale image
segmentation based on pixel classification. Each pixel was assigned differential geometrical
invariants features. The features vector was fed into a Kohonen neural network obtaining a
so-called prototypical feature patterns. A training image was used to train the architecture.
Any image similar with the training image could be then segmented by comparing the feature
pattern representation of each pixel. Thus, supervised labelling was used to classify pixel as
belonging to classes derived from the segmentation process of the training image.
All supervised methods described so far have a major shortcoming: they all require a
training phase using a large set of sample images before run-time. In most of the practical
real-time applications, gathering of sample images is an expensive, laborious and time
consuming process, if not an almost impossible process.

9
Unsupervised techniques
The difference between supervised and unsupervised technique is presented in figure
3. The method is automatically segmenting the image into k sub-regions and then
automatically assigns labels to those regions.

Fig.3 Unsupervised NN-based image segmentation techniques
There are many literature reports for supervised NN-based image segmentation techniques.
Among these, we will only present techniques based on Hopfield NN (HNN), Kohonen Self
Organising Maps (SOM), Constraint Satisfaction NN (CSNN), oscillatory NN (LEGION),
pulse-coupled NN(PCNN).
Neural hypercolumn architecture was used successfully in segmenting radiographic
weld images [30]. A neural hypercolumn consists of stack of nodes that can detect changes
in the spatial intensity gradient from left to right. All pixels in the left-half image are viewed
as inhibitory, whereas pixels on the right are viewed as excitatory. Changes in the intensity
gradient at the boundary are detected by activation of the node. Classical techniques for edge
detection, such as Sobel filters or other edge-detection techniques can be implemented using
such sophisticated artificial hypercolumn architectures.

Hopfield ANN (HNN)
HNN was proposed in 1985 by Hopfield as a way of solving optimisation problems.
In a HNN each neuron is linked to another and weights are symmetrical, i.e. w
ij
=w
ji
, where
w
ij
represent the weight of connection between neuron i and j. The network for the
optimisation application tends to relax into stable states that minimises an energy function of
the form [31,32,33]:
10




N
i
N
j
N
i
i i j i ij
v I v v w E
1 1 1
,
(5)

where N is the number of neurons, v
i
is the output of the i
th
neuron , and I
i
is the external
input for the i
th
neuron term. Hopfield demonstrated that HNN relax into a stable state
tending to minimise its corresponding energy function.
The strategy used by the majority of the authors comprises two steps: first some
means of finding a binary representation for the segmentation solution, so that they can be
mapped into a HNN stable state; and secondly, the definition of the energy function whose
minimisation will lead to an optimum solution to the problem. Such an energy function must
comprise terms for image segmentation constraints E
constraints
or syntax energy i.e. to ensure
that no grey-level or pixel can belong to two classes in the same time, and terms for goodness
of segmentation, E
goodness
or the semantics energy:
goodness s constraint
E E E +
(6)
Amartur et al. [34] used a HNN with winner-take-all (WTA) neurons for
segmentation of MRIs. A grid of Nxk neurons was used, where N is the number of pixels in
the image and k is the predefined number of classes or regions. The use of T2-weighted and
density-weighted MR images of the head allows the network to minimise an energy function
of form:

N
k
N
l
kl kl
V R E
1 1
2
2
1
, (7)
where R
kl
is the distance measure between the k
th
pixel and the centroid of class l and V
kl
is
the output of the k
th
neuron for class l.
A similar approach was taken by Cheng et al. [35] for the segmentation of medical
images. They have used a modified version of HNN with integrated Winner Take All (WTA)
11
learning mechanism called Competitive Hopfield neural network (CHNN). The structure of
CHNN used is independent of the image size being dependent only by the number of
predefined classes and the number of existing grey-levels, as shown in figure 4. In this case
the energy function that needs to be minimised is defined as the mean of squared distance
measures of the grey levels within each class. The Lyapunov-based energy function used is:

n
x
n
y
k
i
yi y xy xi n
y
yi y
v h d v
v h
E
1 1 1
1
1
,
(8)
where n is the number of grey levels in the image, k is the number of predefined classes or
segments, d
xy
is the square of the Euclidian distance measure between the grey-level pairs x
and y and h
x
is the number of pixel at the grey-level x. The performances of CHNN and Hard
C-means Algorithm [35] were compared on both computer simulated, CT and MR images.
The metrics used for comparison were simple and a more appropriate comparison needed to
be done. Furthermore, the computational level of the CHNN is high, and there is no guarantee
that the network does not settle down to local minima.


Fig.4 CHNN proposed by Cheng et al. where n is the number of grey-levels and k is the
number of predefined classes

12
Poli and Valli [36] describe the segmentation as a small set of two-dimensional layers
of neurons, idea taken from the colouring process of geographical maps (figure 5).

Fig.5 Segmentation of a synthetic image (left) using only four colours (center) and the binary
representation with a layer of neurons for each colour used (right)
They proposed a HNN based on that binary representation for the segmentation of X-ray and
MR images. Their semantic energy function includes the sensitivity energy that forces the
network to reveal any changes in the image grey level and robustness energy that is aimed to
reduce the effects of noise and texture. Moreover, an extension of this approach is proposed
for 3D-image segmentation. A very similar approach is described in [24] for the segmentation
of synthetic and CT images.
As a solution to the fact that a single-layer HNN has major limitations, such as
settling into a local minimum, a multi-layer Hopfield neural network (MLHNN) for object
recognition is presented by Young et al. [37]. Their architecture converges to a local
minimum, but that minimum is often equal or very close to the global minimum. An
extended one-layer version of the MLHNN is also presented. However, the computational
overhead induced by those proposed architectures is not presented. Furthermore, the speed of
the convergence process, which is directly dependant on the number and complexity of
computations, is not compared with that of a single-layer HNN but with different types of
13
updating procedures. Therefore, more investigations should be done to prove the robustness
of MLHNN.
A multi-modal image segmentation method was proposed by Rout et al. [38]. Their
method generates a threshold surface by interpolating the image grey levels at points where
the gradient is high (possible edges). A modified HNN was used for the interpolation process.
The performance is compared with the classical method of interpolation with potential
surfaces, where the discrete Laplacian operation is computed for every pixel of the image,
except the edge pixels. Although, the neural network approach has proved to be up to five
times faster than the conventional approach, due to the property of HNN optimisation, a
valuable comparison of goodness of segmentation for both methods was not performed. The
presence of noise was not investigated at all.
Koss et al.[39] investigated the use of HNN for segmentation of CT or MR images of
abdominal organs. The first step to segmentation was to pre-process the images by computing
second-order statistical texture transforms [40]. The results are then input into a HNN. The
HNN structure is presented in figure 6.

Fig.6 HNN structure proposed by Koss et al.; a two-dimensional matrix of neurons: the size
of the rows being the number of predefined classes k, and the size of the columns being the
number of pixels in the image N; an active neuron (i,j) means that the pixel i is assigned to
predefined class j.
14

The energy function proposed in this case is:



+
1
1
]
1

,
_

,
_

+
k
p
N
i
pi pi
N
i
k
p
k
p q
N
i
k
p
pi qi pi
v R v v v A E
1 1
2
1 1 1
2
1
1 , (9)
where R
pi
is the distance of point i from centroid of class p. The first two terms combines the
constraint conditions of the segmentation (the syntactic energy), i.e. each pixel must be
assigned to only one class so that each row must have only one neuron active and the
summation of each row must be one in order that a pixel is assigned 100% to a class. The
semantic energy is defined by the third term that forces the network to find the minimum of
the summation of the distances between the centroids of the predefined classes. In order to
decrease the computation overhead and memory requirements, a WTA scheme is applied so
that the energy equation (9) is transformed like in [34](7):

k
p
N
i
pi pi
v R E
1 1
2

(10)
The performance of this method was proved to be above 90% of the pixels for most of the
organs present in the abdominal cavity. The presence of noise in the original images was still
not investigated. The authors proposed a series of methods of improvement, such as speeding
up the convergence process, reducing the number of neurons [35], and using a continuous-
HNN.

Self-Organising Maps ( SOMs)
Valli et al. [24] proposed an architecture based both on anatomical or biological
knowledge and the physics of image generation. The overall structure of their system is
presented in figure 7. The input image is transformed using image generation physics i.e.
15
enhances information related to signal decay and improves the signal-to-noise ration of the
image. Furthermore, anatomical knowledge is used to generate an initial predefined
segmentation of the original images. The resultant images are then fed into the neural
classifier that performs the segmentation. The modules that incorporates the anatomical and
image generation physics are based on BPNN, while the neural classifier consists of a one-
dimensional Kohonen self-organising map used jointly with a unit labelling algorithm [41].

Fig.7 Segmentation system based on both anatomical knowledge and physical image
generation

A Kohonen self-organising map with a competitive learning algorithm was used by
Reddick et al. [42] to segment MRIs. This method makes use of T1-weighted, T2-weighted,
and proton density-weighted MR images as inputs for the SOM. Thus, there are no spatial
constraints involved. A nine-level grey-scale image was the result of the segmentation. After
segmentation, a three-layer BPNN was trained to classify each pixel according to the tissue
type. In this case, the segmentation was made fully automated, but the classification process
still requires a priori knowledge, used to train the BPNN. Therefore this method could be
16
classified as a hybrid method between supervised and unsupervised image segmentation
techniques. No noise influence was analysed.
A two-step method for segmenting multispectral satellite images was proposed by
Ambroise et al. [43]. The distribution of the pixels to be classified is analysed using a
Probabilistic Self-Organising Map (PSOM) and then Agglomerative Hierarchical Clustering
(AHC) is used. Because AHC cannot be used for clustering millions of objects, PSOM
provides an initial partition of the image, so that AHC can be applied afterwards. The
technique proved to have results that were comparable with classical techniques.
Furthermore, the authors claim that different levels of classification are possible using a
process that needs to be investigated in the future.

Constraint Satisfaction Neural Networks (CSNN).
Lin et al. [44] proposes a new class of neural network for image segmentation
Constraint Satisfaction Neural Networks (CSNN). This method is based on the assumption
that an image segmentation problem can be described as a Constrained Satisfaction Problem
(CSP). The segmentation problem is redefined as the process of assigning each pixel a label
according to certain spatial constraints. A CSNN consists of n x n x m neurons, where n x n is
the image size and m is the number of classes the objects needs to be segmented into (called
labels. Each neuron is connected to all its neighbours (8 connections). Those connections
are the spatial constraints on the segment label of each pixel. Thus, a CSNN comprises a set
of objects, a set of labels (classes) and spatial constraints describing the relationship between
various objects according to neighbour relations. They are updated so that a neuron will
excite other neurons belonging to the same class and inhibit the others. A winner-take-all
scheme is chosen in order to deal with the crossover between segments. As in the standard
learning algorithms, there are two phases: learning phase and categorization phase. Each
17
neuron receives feedback from its own output and excitatory/inhibitory signals from its
neighbours. This method requires that the number of classes (labels) be known a priori. The
proposed method also requires an initial assignment of label probabilities to every pixel using
algorithms such as K-means, ISODATA, fuzzy c-means or Kohonens self-organising map
(SOM).
Certain improvements to the CSNN method were done by Kurugollu et al. [45] using
explicit edge constraints and Kurugollu and Sankur [46] using pyramidal constraints,
speeding up the process of convergence near the edges of the images. Another architecture
based on a CSNN was proposed by Kurugollu and Sankur [47] Multi-Scan constraint
satisfaction neural network (MS-CSNN). The major advantage of this approach is that the
appropriate number of segments is automatically determined by using a modified Zhang-
Modestino [48] cluster validity index.
Due to the large number of neurons n x n x m, this method requires computations that
usually are not suitable for real-time applications. It was applied successfully to CT
(computed tomography) images and MRIs [43,44,45]. Also, there is no literature about the
behaviour of this CSNN in the presence of noise.

A locally excitatory globally inhibitory oscillator network (LEGI ON)
A locally excitatory globally inhibitory oscillator network (LEGION) was proposed
by Terman and Wang [49,50] as a possible computational method for image analysis and
segmentation. The idea under LEGION lies in the biological world and is based on the
temporal correlation theory. Neurons from the visual cortex respond only to stimuli from a
particular part of the visual scene. Moreover, synchronised behaviour had been observed
between spatially separated neurons so that cortical neurons corresponding to a distinct
homogenous area are oscillating in phase, whereas neurons corresponding to different areas
18
are out of phase. LEGION is able to group similar features in an image by achieving fast
synchrony with local excitation. Furthermore, dissimilar features are separated by
desynchrony with global excitation/inhibition.
Chakravarthy et al. [51] proposed a theoretical network of oscillating neurons for
image segmentation purposes - an ANN of complex-valued neurons that exhibit stimulus-
specific oscillations. In their model, the synchronization of neural oscillations is produced by
both co-operation among neurons via excitatory/inhibitory couplings and by hebbian-like
synaptic modification. Therefore, the networks exhibits coherent oscillations according to the
input images, segmentation occurring due to local uniformities in image intensities. They
suggest that special local features such as texture, orientation, could also be used to achieve
synchronisation between neurons.
Shareef et al. [52] used a single layer LEGION-based neural network for
segmentation of medical images. A one-to-one correspondence between image pixels and
neurons is used. Due to the dynamics of LEGION architecture, pixels with similar features
will lead to oscillatory behaviour from their corresponding neurons and their phases will
work in synchronization. Their adaptive scheme for grouping similar features (intensity
contrast of pixels) has proven to work better on CT images rather than MR images. The speed
achieved was as much as ten times better than classical segmentation techniques, i.e.
statistical methods, global thresholding, active contours-snakes, etc. This architecture can be
made more flexible, in the sense that a second layer can be added to process the result from
the first layer, thus improving the segmentation results.

Pulse-coupled neural networks (PCNNs)
Recently, pulse-coupled neural networks (PCNNs) have been used for image
processing, analysis, including image segmentation [53,54]. PCNN is a single layered two-
19
dimensional neural network. It consists of laterally connected pulse-coupled neurons that are
a modification of an Eckhorns neuron model [55].
The architectures proposed, establish a one-to-one correspondence between image
pixels and pulse-coupled neurons. In [53] conditions that lead to the perfect segmentation of
a two-object image (object and background) are derived. None of other image segmentation
techniques described in the present study proposed conditions that guarantee a perfect
segmentation.
Liu et al. [56] proposed an architecture based on both LEGION and PCNN used to
segment range images. In their ANN, each neuron has excitatory lateral connections to the
neighbours and a connection to a global inhibitor. Each neuron has associated a feature vector
consisting of depth, surface normal, mean and Gaussian curvatures. No a priori knowledge is
necessary about the number of regions/segments.
The novelty of PCNN and LEGION and their application to image
processing/segmentation leads to more research and analysis in order to prove their
robustness and applicability.

Conclusions
Neural-network based techniques are used successfully for the segmentation process.
The results presented in the literature entitled these methods to a bright future, though much
more work and research needs to be done. In the vast majority of the papers presented here,
the results of the segmentation process lack general recognised metrics or a general
evaluation framework. Therefore, comparison between methods is still author subjective
matter of opinion. Moreover, one of the best advantages of neural networks their graceful
degradation in the presence of noise- has not been thoroughly investigated.
20
Many methods presented here still require sample data and a priori knowledge (such
as k the number of desired classes or regions to be segmented). This problem can be approach
by using another method of artificial intelligence genetic algorithms (GAs). There are
already reports about GAs used in classical methods like thresholding [57] and edge detection
[58] or texture segmentation [59]. Another approach of removing the necessity of predefined
knowledge was that of using most recent classes of neural networks, such as PCNN and
LEGION. Due to their flexibility and lack of a priori information, they are worth
investigating further.
No general applicable framework or method has been found for image segmentation, even
though some results in that direction are promising.

References
1. S.W. Lawson, Defect Detection in Industrial Radiographic and Ultrasonic Images
(PhD thesis), University of Surrey, Guildford, UK, 1994
2. C. Amza, M. Graves, P. Innocent, J. Knight, Flexible neural network classifier for the
automated detection of bones in chicken breast meat, Proc. International Conference
on Engineering Applications of Neural Networks 17th - 19th of July 2000, Kingston
University, UK, 2000, in press
3. M. Graves, X-ray Machine Vision for on-line Quality Control Food Processing,
University of Cardiff, 1999
4. K.S. Fu, J.K. Mui, A survey on image segmentation, Pattern Recognition, vol.13,
pp.3-16, 1981
5. R.M. Haralick, L.G. Shapiro, Survey, image segmentation technique, Comput. Vision
Graphics Image Process. (CVGIP), vol.29, pp.100-132, 1985
6. P.K. Sahoo, S. Soltani, A.K.C. Wong, Y.C. Chen, A survey of thresholding
techniques, Comput. Vision Graphics Image Process (CVGIP), vol.41, pp. 233-260,
1988
7. N.R. Pal, S.K. Pal, A review on image segmentation techniques, Pattern Recognition,
vol.26, no.9, pp.1277-1294, 1993
8. J.C. Bezdek, L.O. Hall, L.P. Clarke, Review of MR image segmentation techniques
using pattern recognition, Med. Phys., vol.20, no.4, pp. 1033-1048, Jul/Aug, 1993
9. L.P. Clarke, R.P. Velthuizen, M.A. Camacho, J.J. Heine, M. Vaidyanathan, MRI
Segmentation: Methods and Applications, Magnetic Resonance Imaging, vol.13, no.3,
pp.343-368, 1995
10. J.M. Zurada, Introduction to Artificial Neural Systems, West Publishing Company,
USA, 1992
11. S.Y. Kung, Digital Neural Networks, PTR Prentice Hall, Englewood Cliffs, New
Jersey, 1993
12. S. Haykin, Neural Networks: A comprehensive foundation, Macmillan College
Publishing Company, Inc., USA, 1994
21
13. W.E. Blanz, S.L. Gish, A connectionist classifier architecture applied to image
segmentation, Proc. 10
th
ICPR, pp.272-277, 1990
14. S.W. Lawson, G.A. Parker, Intelligent Segmentation of industrial radiographic images
using neural networks, in Proc. of SPIE, vol.2347, pp.245-255, 1994
15. W.P. de Waard, Neural techniques and postal code detection, Pattern Recognition
Letters, vol.15, pp. 199-205, 1994
16. R.H. Silverman, A.S. Noetzel, Image processing and pattern recognition in
ultrasonograms by backpropagation, Neural Networks, vol.3, pp.593-603, 1990
17. M. Ozkan, H.G. Sprenkels, B.M. Dawant, Multi-spectral resonance image
segmentation using neural networks, in Proc. IJCNN 90, vol.1, pp.429-437, San
Diego, June 1990
18. K. Oshio, M. Singh, Automatic segmentation of magnetic resonance head images
using neural nets, Proc. IEEE Nucl. Sci. Symp. And Medical Imaging Conf., pp.1439-
1434, 1990
19. M. Ozkan, B.M. Dawant, R.J. Maciunas, Neural-Network-Based Segmentation of
Multi-Modal Medical Images: A comparative and Prospective Study, IEEE Trans. On
Medical Imaging, vol.12, no.3, pp.534-544, September 1993
20. U. Raff, A.L. Scherzinger, P.F. Vargas, J.H. Simon, Quantitation of grey matter,
white matter, and cerebrospinal fluid from spin-echo magnetic resonance images
using an artificial neural network technique, Med.Phys., vol.21, no.12, pp.1933-1942,
December 1994
21. N. Babaguchi, K. Yamada, K. Kise, Connectionist model binarization, Proc. 10
th

ICPR, pp.51-56, 1990
22. L.O. Hall, A.M. Bensaid, L.P. Clarke, R.P. Velthuizen, M.S. Silbiger, J.C. Bezdek, A
comparison of neural networks and fuzzy clustering techniques in segmenting
magnetic resonance images of the brain, IEEE Trans. Neural Network, vol. 3, no.5,
pp.672-682, 1992
23. Coppini, G., Poli, R., Legitimo, R., De Dominicis, R., Valli, G., A neural network
system for detecting lung nodules in chest radiograms. In Computer Assisted
Radiology, CAR93, pp. 594-599, Springer-Verlag, Berlin, 1993
24. G. Valli, R. Poli, S. Cagnoni, G. Coppini, Neural Networks and Priori Knowledge
Help the Segmentation of Medical Images, Journal of Computing and Information
Technology- CIT 6, vol.2, pp.117-133, 1998
25. T. Wang, X. Zhuang, X. Xing, Robust segmentation of noisy images using neural
network model, Image and Vision Computing, vol.10, pp.233-240, 1992
26. K. Fukanage, Introduction to Statistical Pattern Recognition, New York: Academic,
pp.260-267, 1972
27. K. Shiranita, K. Hayashi, A. Otsubo, T. Miyajima, R. Takiyama, Determination of
Meat Quality by Image Processing and Neural Network Techniques, Proc. Of
FUZZIEEE2000, Vol. 2, pp516-521, San Antonio, Texas, May 2000
28. S. Haring, M.A. Viergever, J.N. Kok, Kohonen networks for multiscale image
segmentation, Image and Vision Computing, vol.12, no.6, pp.339-344, July/August
1994
29. T. Kohonen, Self-Organization and Associative Memory, Third Edition, Springer-
Verlag Berlin, 1989
30. A. Gaillard, D.C. Wunsch II, R.A. Escobedo, Neural hypercolumn architecture for the
preprocessing of radiographic weld images, SPIE, vol.1294 Applications of Artificial
Neural Networks, pp.378-388, 1990
31. J.J. Hopfield, Neural networks and physical systems with emergent collective
computational abilities, Proc. Natl.Acad.Sci., USA, vol.79, pp.2554-2558, April, 1982
22
32. J.J. Hopfield, Neurons with graded response have collective computational properties
like those of two-state neurons, Biophysics: Proc.Natl.Acad.Sci., USA, vol.81,
pp.3088-3092, May, 1984
33. J.J. Hopfield, D.W. Tank, Neural Computation of Decisions in Optimisation
Problems, Biol.Cybern., vol.52, pp.141-152, 1985
34. S.C. Amartur, D. Piraino, Y. Takefuji, Optimization Neural Networks for the
Segmentation of Magnetic Resonance Images, IEEE Transactions on Medical
Imaging, vol.11, no.2, pp.215-220, June 1992
35. K.S. Cheng, J.S. Lin, C.W. Mao, The application of competitive Hopfield Neural
Network to medical image segmentation, IEEE Transactions on Medical Imaging,
vol.15, no.4, pp.560-567, August 1996
36. R. Poli, G. Valli, Hopfield neural networks for the optimum segmentation of medical
images, Handbook of Neural Computation, Oxford University Press, chapter.G5.5,
pp.1-10, 1997
37. S.S. Young, P.D. Scott, N.M. Nasrabadi, Object Recognition using Multilayer
Hopfield Neural Network, IEEE Transactions on Image Processing, vol.6, no.3, pp.
357-371, March 1997
38. S. Rout, P. Srivastava, J. Majumdar, Multi-modal image segmentation using a
modified Hopfield Neural Network, Pattern Recognition, vol.31, no.6, pp.743-750,
1998
39. J.E. Koss, F.D. Newman, T.K. Johnson, D.L. Kirch, Abdominal organ segmentation
using texture transforms and a Hopfield neural network, IEEE Transactions on
Medical Imaging, vol.18, no.7, pp.640-648, July 1999
40. R.M. Haralick, K. Shanmugam, I. Dinstein, Textural features for image classification,
IEEE Trans. Syst., Man, Cybern., vol. SMC-3, no.6, pp.610-621, 1973
41. D.H. Wolpert, Stacked generalisation, Neural Networks, vol.5, pp.241-259, 1992
42. W.E. Reddick, J.O. Glass, E.N. Cook, T.D. Elkin, R.J. Deaton, Automated
Segmentation and Classification of Multispectral Magnetic Resonance Images of
Brain Using Artificial Neural Networks, IEEE Transactions on Medical Imaging,
vol.16, no.6, pp.911-918, December 1997
43. C. Ambroise, G. Seze, F. Badran, S. Thiria, Hierarchical clustering of self-organizing
maps for cloud classification, Neurocomputing, vol.30, pp.47-52, 2000
44. W. Lin, E. Chen-Kuo, C.T. Chen, Constraint satisfaction neural networks for image
segmentation, Pattern Recognition, vol.25, no.7, pp.679-693, 1992
45. F. Kurugollu, S. Birecik, M. Sezgin, B. Sankur, Image segmentation based on
boundary constraint neural network. In Proc. Third International Workshop on Image
and Signal Processing, IWISIP96, Manchester, UK, pp.353-256, 1996
46. F. Kurugollu, B. Sankur, Colour cell image segmentation using pyramidal constraint
satisfaction neural network. In IAPR Workshop on Machine Vision Applications
MVA98, 17-19 November 1998, Makuri, Chiba, Japan, 1998
47. F. Kurugollu, B. Sankur, Image segmentation based on multi-scan constraint
satisfaction neural network, Pattern Recognition Letters, no.20, pp.1553-1563, 1999
48. J. Zhang, J.W. Modestino, Model-fitting approach to cluster validation with
application to stochastic model-based image segmentation, IEEE Trans. PAMI 12, pp.
1009-1017, 1990
49. D. Terman, and D.L. Wang, Global competition and local cooperation in a network of
neural oscillators, Physics D, vol. 81, pp.148-176, 1995
50. D.L. Wang, D. Terman, Image segmentation based on oscillatory correlation, Neural
Computing, vol.9, pp.805-836, 1997
23
51. S. Chakravarthy, V. Ramamurti, J. Ghosh, A network of oscillating neurons for image
segmentation, Intelligent Engineering through ANN, vol.5, Dagli, Akay, Chen
Fernanded and Ghosh (editors), ASME Press, Nov. 1995
52. N. Shareef, D.L. Wang, R. Yagel, Segmentation of Medical Images using Legion,
IEEE Transactions on Medical Imaging, vol.18, no.1, pp.74-91, January 1999
53. H.S. Ranganath, G. Kuntimad, Object detection using pulse-coupled neural networks,
IEEE Transactions on Neural Networks, vol.10, no.3, pp. 615-620, May 1999
54. G. Kuntimad, H.S. Ranganath, Perfect Image Segmentation Using Pulse Coupled
Neural Networks, IEEE Transactions on Neural Networks, vol.10, no.3, pp/.591-598,
May 1999
55. R. Eckhorn, H.J. Reitboeck, M. Arndt, and P. Dicke, Feature linking via
synchronization among distributed assemblies: Simulations of results from cat visual
cortex, Neural Computing, vol.2, pp.293-307, 1990
56. X. Liu, D.L. Wang, Range Image Segmentation Using a Relaxation Oscillator
Network, IEEE Transactions on Neural Networks, vol.19, no.3, pp. 564-573, May
1999
57. P.-Y. Yin, A fast scheme for optimal thresholding using genetic algorithms, Signal
Processing 72, pp.85-95, 1999
58. M. Gudmundsson, E.A. El-Kwae, M.R. Kabuka, Edge detection in medical images
using a genetic algorithm, IEEE Transactions on Medical Imaging, vol.17, no.3,
pp.469-474, June 1998
59. M. Yoshimura, S. Oe, Evolutionary segmentation of texture image using genetic
algorithms towards automatic decision of optimum number of segmentation areas,
Pattern Recognition, 32, pp.2041-2054, 1999

Acknowledgements
The author wishes to thank Prof. Peter Innocent for his invaluable input and ideas to improve
the paper. Special thanks are due to Prof. Jeffrey Knight for his suggestions and support.

Glossary of Terms and Abbreviations
used:
NN = neural network
MRI = magnetic resonance image
CT = computer tomography image
ANN = artificial neural network
FFNN = feed forward neural network
BP = back-propagation learning algorithm
BPNN = back-propagation-based neural
network
MLFF = multi-layer feed forward neural
network
MLP = multi-layer perceptron
HNN = Hopfield neural network
WTA = winner-takes all learning
mechanism
CHNN = competitive Hopfield neural
network
MLHNN = multi-layer Hopfield neural
network
SOM = self-organising map (Kohonen)
PSOM = probabilistic self-organising map
AHC = agglomerative hierarchical
clustering
CSNN = constraint satisfaction neural
network
CSP = constraint satisfaction problem
MS-CSNN = multi-scan constraint
satisfaction neural network
LEGION = locally excitatory globally
inhibitory oscillator network
PCNN = pulse-coupled neural network
GA = genetic algorithm