CHAPTER-3 IMAGE SEGMENTATION...
Transcript of CHAPTER-3 IMAGE SEGMENTATION...
�
�
���
�
CHAPTER-3
IMAGE SEGMENTATION TECHNIQUES
3.1 Introduction to Digital Image Processing
Digital image processing involves processing and manipulation of image
for better understanding and enhanced visual perception. There are two primary
motivations behind the emergence of image processing: enhancement of pictorial
information for better human interpretation and processing of scene data for
autonomous machine perception. The term “digital image processing” refers to
processing of digital images by means of digital computers [36]. The term ‘digital
computers’ in this definition not only denotes general purpose digital computers
such as a desktop Personal Computer(PC) or a workstation but also systems that
ranges from simple digital circuits to advanced parallel computers.
The history of image processing dates back to 1920 when Harry G.
Bartholomew and Maynard D. McFarlane invented the Bartlane Cable Picture
Transmission system which transmits images over cable lines across the Atlantic.
Initially only 5 gray levels were used to code the images which increased to 15 in
1929. The Bartlane Cable system cannot be considered as a pure digital image
processing system because analog instruments were used and absence of the
primary concepts used in digital computers. The digital computers began its
journey in the 1940s when John von Neumann put forward the principal concepts
of digital computers. A series of activities mostly comprising of development of
high-level language and invention of integrated circuits(IC) led to the formation of
�
true digital computers. It was only in the early 1960s that the digital computers
became powerful enough to handle the
example of digital image processing is the processing of pictures of the moon
transmitted by the Ranger 7 s
place in the field of medical science, remote sen
physical sciences, archeology, military defen
The first step in digital image processing system is to acquire the images
and store them in storage device. The acquisition of image is done
various types of light sensors.
Metal Oxide Semiconductor
acquisition. Now a day, very high resolution cameras are available. The images
created are the result of illumination and reflection
function of light can be presented by a continuous
two-dimensional light intensity function
coordinates and the value of
the gray level of the image at that point. The function
approximated by equally spaced samples arranged in the form of an
as shown below [36]
���
true digital computers. It was only in the early 1960s that the digital computers
became powerful enough to handle the intricacies of digital image. The earliest
example of digital image processing is the processing of pictures of the moon
transmitted by the Ranger 7 spacecraft. Gradually digital image processing finds its
place in the field of medical science, remote sensing, biological application,
physical sciences, archeology, military defense and numerous other fields.
first step in digital image processing system is to acquire the images
and store them in storage device. The acquisition of image is done with the help of
various types of light sensors. Charged-coupled device (CCD) and Complementary
Metal Oxide Semiconductor (CMOS) sensor devices are mostly used in image
acquisition. Now a day, very high resolution cameras are available. The images
of illumination and reflection of the source light. The
function of light can be presented by a continuous function����� ��. An image is a
dimensional light intensity function ���� ����where x and y are spatial
coordinates and the value of f at any point (x, y) is proportional to the brightness or
the gray level of the image at that point. The function ���� ��approximated by equally spaced samples arranged in the form of an �
�
true digital computers. It was only in the early 1960s that the digital computers
The earliest
example of digital image processing is the processing of pictures of the moon
pacecraft. Gradually digital image processing finds its
sing, biological application,
se and numerous other fields.
first step in digital image processing system is to acquire the images
with the help of
and Complementary
sensor devices are mostly used in image
acquisition. Now a day, very high resolution cameras are available. The images
of the source light. The
An image is a
where x and y are spatial
at any point (x, y) is proportional to the brightness or
� ��� can be
�array
�
�
���
The right side of the above equation is by definition a digital image [36].
The transformation of the continuous analog light function into the discrete matrix
representation is called sampling or quantization.
Figure 3.1: Digitization of continuous image.
The size of an image in computer memory is given by the function
� ���where N and M are the row and column and L is discrete gray levels
allowed for each pixel. L is typically an integer power of 2. The images are store in
computer in different formats each having some distinct characteristics. Some of
the popular and most commonly used formats are TIFF, JPEG, BMP, GIF, PNG
etc. Other image file formats are used to a lesser extent; these formats are often
proprietary, such as Adobe Photoshop .psd files. It is also helpful to understand the
common image file formats of digital images, how these file formats differ, and
what their recommended use is [37].
3.2 Components of Digital image processing systems
A digital image processing system consists of several components, both
hardware and software. A simple personal computer can be used to carry out the
������� ��
�
�
���
�
image processing operations, though there are dedicated systems for image
processing. The dedicated systems are mostly related to specific applications and
industrial process such as microscopic image analysis system or printed circuit
board (PCB) testing system. Although large-scale image processing systems are
still being sold for massive imaging applications, such as processing of satellite
images, the trend continues toward miniaturizing and blending of general-purpose
small computers with specialized image processing hardware. Figure 3.2 depicts
the principal components of a typical digital image processing system. Each of the
components is discussed below:
Light sensors: It is a physical device which is sensitive to the energy
radiated by the object, whose image we want to capture. The second, called a
digitizer, is a device for converting the output of the physical sensing device into
digital form.
Image processing hardware: It usually consists of the digitizer just
mentioned and hardware that performs other primitive operations, such as an
arithmetic logic unit (ALU), which performs arithmetic and logical operations in
parallel on entire images. An example of how an ALU is used is in averaging
images as quickly as they are digitized, is the noise reduction process.
The computer: The computer used in an image processing system ranges
from a PC to a supercomputer. In dedicated applications, sometimes specially
designed computers are used to achieve a required level of performance. For a
general-purpose image processing systems almost any well-equipped PC-type
machine is suitable for offline image processing tasks.
�
�
���
�
Software: Software for image processing consists of specialized modules
that perform specific tasks. A well-designed package also includes the capability
for the user to write code that, as a minimum, utilizes the specialized modules.
More sophisticated software packages allow the integration of those modules and
general-purpose software commands from at least one computer language. e.g.
Champ Software, MATLAB Image Processing Tool etc.
Mass storage: Storage capability is a must in any image processing
applications. An image of size 1024*1024 pixels, in which the intensity of each
pixel is an 8-bit quantity, requires one megabyte of storage space if the image is
not compressed. Thus dealing with thousands, or even millions of images,
providing adequate storage in an image processing system can be a challenging in
real time operation.
Image displays: The image displays in use today are mainly colour
monitors. Monitors are driven by the outputs of image and graphics display cards
that are an integral part of the computer system. Seldom are there requirements for
image display applications that cannot be met by display cards available
commercially as part of the computer system
Hardcopy devices: They include laser printers, film cameras, heat-
sensitive devices, inkjet units, and digital units, such as optical and CD-ROM
disks. Film provides the highest possible resolution, but paper is the obvious
medium of choice for written material. For presentations, images are displayed on
film transparencies or in a digital medium if image projection equipment is used.
The latter approach is gaining acceptance as the standard for image presentations.
�
�
���
�
Networking: It is almost a default function in any computer system in use
today. Because of the large amount of data inherent in image processing
applications, the key consideration in image transmission is the required
bandwidth. In dedicated networks, this is not a problem, but communications with
remote sites via the Internet are of much concern. Fortunately, this situation is
improving quickly as a result of optical fiber and other broadband technologies 3G
network.
Figure 3.2: Components of Digital image processing system
3.3 Basic Steps in Digital Image Processing
An image processing system processes the image through several stages to
achieve the appropriate output. The efficiency of the result depends on every stage
of processing. The basic steps in digital image processing are described below.
�
�
���
�
Image acquisition: Image acquisition is the first step of digital image
processing. It is the process of capturing the images of some objects and
transforming them into suitable digital formats for storing in the computer. Light
sensors are used to capture the image. Typically charged coupled device (CCD)
and complementary metal oxide semiconductor devices (CMOS) are used as light
sensors. The CCD or CMOS sensors captures analog signal of lights which are not
appropriate for computer processing. Another device, called Digitizer, transforms
the analog signal to digital form. The transformation process is called sampling and
quantization [38] [39].
Enhancement and Restoration: Various factors such as visibility,
brightness and contrast of grey level of pixels, amount of noise present are
responsible for the quality of an image. Before any actual processing, images need
to be in a state which is favourable for processing. There are two tasks:
enhancement and restoration. The idea behind enhancement techniques is to bring
out detail that is obscured, or simply to highlight certain features of interest in an
image. A familiar example of enhancement is when we increase the contrast of an
image because it looks better. Enhancement is a very subjective area of image
processing. Image restoration is an area that deals with improving the appearance
of an image. However, unlike enhancement, which is subjective, image restoration
is objective. The restoration techniques used must be based on mathematical or
probabilistic models of image degradation [40].
Morphological Processing: Morphological processing deals with tools for
extracting image components that are useful in the representation and description
of shape. At first, morphological operations were defined for only binary image
�
�
���
�
that is an image pixel value either 0 or 1. Later, several morphological techniques
have evolved which can handle gray level image also [40]. Morphological
operations apply a structuring element to an input image, creating an output image
of the same size. In a morphological operation, the value of each pixel in the output
image is based on a comparison of the corresponding pixel in the input image with
its neighbors [41]. By choosing the size and shape of the neighborhood, one can
construct a morphological operation that is sensitive to any specific shapes in the
input image.
Segmentation: It is the process of dividing an image into its constituent
parts. The goal of segmentation is to partition the object of interest according to the
requirement of specific application. Image segmentation is discussed in details in
Section 3.4.
Representation and Description: Representation and description always
follow segmentation, which is usually raw pixel data [36]. There are mainly two
ways of representation: boundary representation, which is appropriate when the
focus is on external shape characteristics, such as corners and inflections, and
regional representation is appropriate when the focus is on internal properties, such
as texture or skeletal shape. Description, also called feature selection, deals with
extracting attributes that result in some quantitative information of interest. They
are basic for differentiating one class of objects from another.
Object Recognition: It is the process that assigns a label to an object based
on its descriptors. For any object in an image, there are many 'features' which are
interesting points on the object that can be extracted to provide a "feature"
description of the object. This description extracted from a training image can then
�
�
���
�
be used to identify the object when attempting to locate the object in a test image
containing many other objects as well [42] [43].
Figure 3.3depicts the flow of these basic steps in digital image processing.
Figure 3.3: Steps in Digital Image Processing
3.4 Image Segmentation
Image segmentation is the most basic and important part of image
processing which segments an image into meaningful areas according to some
characteristics such as gray level, spectrum, texture, colour, and so on. The goal of
image segmentation is to partition an image into a set of disjoint regions with
uniform and homogeneous attributes such as intensity, color, tone or texture etc
[44]. More precisely, image segmentation is the process of assigning a label to
every pixel in an image such that pixels with the same label share certain visual
characteristics.
Segmentation of nontrivial images is one of the most difficult tasks in
image processing [36]. The success of the subsequent image processing tasks
�
�
���
�
depends on the robustness of the segmentation process. So a considerable amount
of care should be taken to improve the probability of the rugged segmentation.
There is no generalized theory of image segmentation. As a consequence, no single
standard method of image segmentation has emerged. Rather, there are a collection
of ad hoc methods that have received some degree of popularity. Most of the image
segmentation algorithms are based on two basic properties of intensity values:
discontinuity and similarity. The ‘discontinuity’ approach is to segment the image
based on abrupt change in intensity, such as edged in the images. The second
approach is based on some pre defined criteria upon which images are partitioned.
Thresholding, Region growing are some methods which fall into this category.
When designing the segmentation algorithm, most of the image processing
designer follows some paradigms such as [36]:
i. Regions of the image segmentation should be uniform and homogeneous
with respect to some characteristic such as gray tone or texture.
ii. Region interiors should be simple and without many small holes.
iii. Adjacent regions of segmentation should have significantly different
values with respect to the characteristic on which they are uniform.
iv. Boundaries of each segment should be simple, not ragged, and must be
spatially accurate.
Mathematically, image segmentation can be defined as a spatial region R
divided into n subregions, R1, R2,….Rn such that
(a) � � � ��������
(b) ������is a connected set, � � �� �� �� � �
�
�
���
�
(c) � � ����� � ���for all i and j, � � � (d) �� �� � � !��"#�� � ���� ��
(e) �� �� $ �)=%&�'!��"#�(���()�(*+�,�#+-�"�.� � and �
Here, �� �� is a logical predicate defined over the points in set � and � is
a null set. From the above equations, we can say that the segmentation is complete
and every pixel in a region is connected in some way. Also, they indicate that the
regions must be disjoint and satisfy some properties of the pixels. There are several
methods of image segmentation [36] some which are discussed in the next section.
3.4.1 Thresholding
The simplest method of image segmentation is called the thresholding
method. This method is based on a clip-level (or a threshold value) to turn a gray-
scale image into a binary image. The key of this method is to select the threshold
value (or values when multiple-levels are selected). Several popular methods used
in industry are: the maximum entropy method, [36] Otsu's method (maximum
variance) [44]. k-means clustering [44] can also be used in this regard.
Thresholding is based on the fact that different objects in an image have different
level of luminance or colour value. Luminance thresholding is of two types: bilevel
luminance thresholding, that is, images having only luminance level such as a
printed page having black letters on white background; and multilevel luminance
thresholding that is an image with multiple value of luminance [45]. Colour
thresholding is based on a nonstandard colour component, loosely called intensity.
The basic principle of thresholding is to set a luminance or colour intensity
threshold, T and comparing each pixel (x, y) of an image f(x,y) with T. If at the
�
�
���
�
point (x,y), f(x,y)>T, then the point is called an object point; otherwise called a
background point. The segmented image g(x,y) is defined as [36]
-��� �� � /�������������� �� 0 �1����������� �� 2 � 3
When T is constant over the entire image, the process is called global
thresholding. When T changes over an image it is called variable thresholding.
There is also a third kind called local or regional thresholding where the value of
T at any point (x,y) depends on the properties of its neighbourhood. If T depends on
spatial coordinates (x,y) themselves, then variable thresholding is called adaptive
thresholding.
The presence of noise effects the outcome thresholding segmentation.
Image smoothing and thresholding with edge detection are two common methods
to nullify the effect of noise on thresholding. In statistics and image processing, to
smooth a data set is to create an approximating function that attempts to capture
important patterns in the data, while leaving out noise or other fine-scale
structures/rapid phenomena. In smoothing, the data points of a signal are modified
so that number of individual points (presumably because of noise) is reduced, and
points that are lower than the adjacent points are increased leading to a smoother
signal [46]. Edge detection is a fundamental tool in image processing, machine
vision and computer vision, particularly in the areas of feature detection and
feature extraction [39].
N. Otsu [47] proposed a global thresholding method where the weighted
sum of within-class variances of the foreground and background pixels is
�
�
���
�
minimized to establish an optimum threshold. A 2D gray-level intensity function
f(x, y) whose value is the gray-level can generally show the characteristics of an
image. If it ranges from 0 to L-1, where L is the number of distinct gray-levels, and
the number of pixels with gray-level � in �� be the total number of pixels in a given
image, the probability of occurrence of gray-level ��is defined as [47]
nnp ii /=
The average gray-level of the entire image is computed as [47]
�−
==
1
0
L
iiT ipµ
For single thresholding, the pixels of an image will be divided into two
classes, C1= {0, 1,…, t} and C2 = {t+1, t+2,…, L-1}, where t is the threshold value.
C1 and C2 correspond to the objects of interest and the background. The
probabilities of the two classes are obtained as [47]
��−
+==
==1
12
01 )(,)(
L
tii
t
ii ptpt ωω
The means of the two classes can be computed as [47]
��−
+====
1
122
011 )(/)(,)(/)(
L
tii
t
ii tipttipt ωµωµ
Using discriminant analysis, Otsu [47] showed that the optimal threshold t*
can be determined by maximizing the between-class variance; defined as [47]
�
�
���
�
)}({ 2
0
* tMaxArgt BLt
σ<≤
=
Where the between-class variance, 45��6 �,���is defined as [47]
222
211
2 ))()(())()(()( TTB ttttt µµωµµωσ −+−=
3.4.2 Region based methods
There are three primary methods for region based segmentation of image.
Out of these, region growing is the simplest one. Region growing methods involve
grouping of neighboring pixels of similar amplitude together to form a segmented
region. However, some constraints must be placed on the growth pattern to achieve
acceptable results [39]. Brice et al. [48] proposed a region-growing method based
on a set of simple growth rules. In the first stage of the process, pairs of quantized
pixels are combined together in groups called atomic regions if they are of the
same amplitude. Two heuristic rules are invoked to dissolve weak boundaries
between atomic boundaries. Referring to Figure 3.4, let R1 and R2 be two adjacent
regions with perimeters P1 and P2, respectively, which have previously been
merged. After the initial stages of region growing, a region may contain previously
merged subregions of different amplitude values. Also, let C denote the length of
the common boundary and let D represent the length of that portion of C for which
the amplitude difference Y across the boundary is smaller than a significance
factor. The regions R1 and R2 are then merged if
7��89�� 9�: 0 ;6
�
Figure 3
;6 is a constant typically set to
method provides reasonably accurate segmentation of simple scenes with few
objects and little texture
Yakimovsky [49] has attempted to improve the region
establishing merging constraints based on estimated Bayesian probability densities
of feature measurements of each region.
The second method based on region is
merge image segmentation techniques are based on a quad tree data representation
whereby a square image segment is broken (split) into four quadrants if the original
image segment is non-uniform in attribute. If four neighboring squares are found to
be uniform, they are replaced (merge) by a single square composed of the four
adjacent squares [50]. Splitting and merging attempts to divide an image into
uniform regions. The basic representational structure is pyramidal, i.e. a square
region of size < =< at one le
below it. Usually the algorithm starts from the initial assumption that the entire
image is a single region, and then computes the homogeneity criterion to see if it is
TRUE. If FALSE, then the square regi
���
Figure 3.4: Region growing geometry
is a constant typically set to � �> . The Brice et al. [48] region growing
method provides reasonably accurate segmentation of simple scenes with few
objects and little texture. But it did not perform well on more complex scenes.
Yakimovsky [49] has attempted to improve the region-growing concept by
erging constraints based on estimated Bayesian probability densities
of feature measurements of each region.
The second method based on region is split and merge method
image segmentation techniques are based on a quad tree data representation
whereby a square image segment is broken (split) into four quadrants if the original
uniform in attribute. If four neighboring squares are found to
they are replaced (merge) by a single square composed of the four
Splitting and merging attempts to divide an image into
uniform regions. The basic representational structure is pyramidal, i.e. a square
at one level of a pyramid having 4 sub-regions of
below it. Usually the algorithm starts from the initial assumption that the entire
image is a single region, and then computes the homogeneity criterion to see if it is
TRUE. If FALSE, then the square region is split into the four smaller regions. This
�
region growing
method provides reasonably accurate segmentation of simple scenes with few
perform well on more complex scenes.
growing concept by
erging constraints based on estimated Bayesian probability densities
split and merge method. Split and
image segmentation techniques are based on a quad tree data representation
whereby a square image segment is broken (split) into four quadrants if the original
uniform in attribute. If four neighboring squares are found to
they are replaced (merge) by a single square composed of the four
Splitting and merging attempts to divide an image into
uniform regions. The basic representational structure is pyramidal, i.e. a square
regions of same size
below it. Usually the algorithm starts from the initial assumption that the entire
image is a single region, and then computes the homogeneity criterion to see if it is
into the four smaller regions. This
�
�
���
�
process is then repeated on each of the sub-regions until no further splitting is
necessary. The process terminates when no further merges are possible. These
small square regions are then merged if they are similar to give larger irregular
regions. However, the problem with this method is that any two regions may be
merged if adjacent and if the larger region satisfies the homogeneity criteria, but
regions which are adjacent in image space may have different parents or be at
different levels in the pyramidal structure.
Figure 3.5: Partitioned image and corresponding quad tree
The third method based on region is watershed method. The basic concept
of watershed method is to visualize the image in three dimensions: two spatial
coordinates versus intensity. In such topographic and hydrologic interpretation,
three types of points are considered: points belonging to a regional minimum,
points where a drop of water, if placed on any of these points would fall certainly
to a local minimum, and points where water equally likely to fall to more than one
such minimum. The accumulation of water in the vicinity of a local minimum is
called a catchment basin. All points that drain into a common catchment basin are
part of the same watershed. Hence a region surrounded by ridge is called a valley
[51]. A ridge is the loci of maximum gradient of the altitude surface. There are two
�
�
���
�
basic algorithmic approaches to the computation of the watershed of an image:
rainfall and flooding. In the rainfall approach, local minima are found throughout
the image. Each local minimum is given a unique tag. Adjacent local minima are
combined with a unique tag. Next, a conceptual water drop is placed at each
untagged pixel. The drop moves to its lower-amplitude neighbor until it reaches a
tagged pixel, where it assumes the tag value. Figure 3.6(a) illustrates a section of a
digital image encompassing a watershed in which the local minimum pixel is
represented as black square and the dashed line represents the path of a water drop
to the local minimum [51].
In the flooding approach, conceptual single pixels are placed at each local
minimum. A sea of pixel proceeds to fill up the neighbouring are of the pixel
located at the local minima. If there is over flow of pixels, the extra pixels are
discarded. Figure 3.6(b) shows a profile of the filling process of a catchment basin
[51].
(a) (b)
Figure 3.6: (a) Rainfall watershed and (b) Flooding watershed
�
�
���
�
3.5 Image segmentation based on fuzzy clustering techniques
Fuzzy clustering is known for its ability to handle uncertainties artificial
intelligence. While the application of fuzzy sets in cluster analysis and classifier
design was in progress, application of fuzzy clustering in image processing was
evolving in parallel to the aforesaid development [52]. This development was
based on the realization that many of the basic image processing concepts, such as
edge or region boundary do not lend themselves well to precise definition. A gray
tone image possesses ambiguity within pixels due to multi level brightness in the
image. Conventional image segmentation methods divide an image into
meaningful parts based on some criteria. The partition involves only two-valued
logic. A pixel can be in one of the segmented portions or it cannot. Higher level of
processing yields a great amount of erroneous results due to this rigidity. A
classical example of this drawback is the skeleton extraction of a region through
medial axis transformation. The medial axis transformation of a region in a gray
tone image is defined based on its boundary. The boundary of gray tone is well
defined and therefore the medial axis transformation is likely to give erroneous
result if we compute it from crisp segmented image. So, it is convenient to avoid
fixed decisions regarding the segmentation criteria by incorporating flexibility of
fuzzy clustering.
There are several fuzzy clustering algorithms which are used in image
segmentation; Fuzzy c-Means (FCM) algorithm is the most widely used one. The
FCM algorithm is based on an objective function described in Section 2.5.2 in the
second chapter. The objective function is very suitable to define the relationship
�
�
���
�
among the pixels in an image. The pixels are considered as data points and the
fuzzy classification matrix denotes the memberships of pixels in clusters. FCM
algorithm can be used to segment both gray level and colour images. Also, another
facility is that we can set the number of clusters that is the number of segmented
parts in advance. A deviation from the expected value might help us to find any
fault in the segmentation process. The objective function of FCM algorithm can be
modified and extended as the requirements change. The adaptability of the
objective function makes FCM suitable for different kind of images. Segmentation
of any image needs to be validated after completion. A typical and well practiced
mean is visual observation of the images by the experts of the concerned field.
Mathematical heuristics are available for validating the segmentation some of
which are complex and computationally expensive. FCM has some validity
measures which can be incorporated within the algorithm and thus it minimizes the
computational effort as well as yields satisfactory results. In literature, we can find
a numerous application FCM algorithm in image segmentation.
Trivedi et al. [53] proposed a fuzzy set theoretic image segmentation
algorithm for aerial image segmentation. The method is based on region growing
principal using a pyramid data structure. The algorithm is hierarchical in nature.
Segmentation of the images at a particular level of processing is done by FCM
algorithm. In multi level segmentation, level � regions are considered
homogeneous when image elements have largest cluster membership values of
greater than a prescribed threshold. If homogeneity test fails, regions are split to
form next level of regions which are again subjected to FCM algorithm. Hall et al.
[54] applied FCM with neural network to segment MRI images. Huntsberger et al.
�
�
���
�
[55] applied FCM in image segmentation with refined iterative manner. Keller et
al. [56] used a modified version of FCM in image segmentation. The cluster
centers are updated using the FCM formula, but the new membership values for
each point are calculated using an S-type function based feature value of each point
and the fuzzy means. Backer et al. [57] developed a general strategy of clustering
based fuzzy approach. The set of samples is at first divided into * disjoint sets. The
membership of each point in the initialized clusters is based on some affinity
measures such as distance concept, neighbourhood concept. Three criterion
functions are used based on fuzzy concept. This strategy is very favourable for
image segmentation. Multispectral images are also processed with FCM algorithm
[56] [58] [59].
Lung et al. [60] have developed generalized Spatial Fuzzy C-Means
Clustering algorithm (GSFCM) for brain MRI segmentation. GSFCM utilizes
given pixel attributes and spatial local information weighted equally to neighbours
based on their distance attributes. Results have shown that GSFCM outperforms
conventional FCM. Zhou et al. [61] have presented a mean shift based FCM for
the extraction of skin lesions. They have proposed a mean shift based fuzzy c-
means objective function that is a mean field term is incorporated in the standard
FCM objective function. Experimental results have shown that their algorithm is
capable of extracting skin lesion borders proficiently. Du et al. [62] has proposed
an enhanced segmentation technique that applies sigma filter to every
neighbouring pixel of targets. Visual and quantitative assessment has revealed that
the proposed method works better than the original FCM. Sudhavani et al. [63]
presented modified fuzzy C-Means clustering algorithm for lip image
�
�
���
�
segmentation. Kannan et al. [64] have developed an efficient fuzzy segmentation
algorithm for breast magnetic resonance imaging data. They have acquired an
objective function of FCM called Kernel Induced FCM which is based on hyper
tangent function which is in turn based on two functions namely kernel, hyper
tangent and Langrangian multipliers. Yang et al. [65] describes a modified FCM
objective function based on presence of noise. Guo et al. [66] proposed a
segmentation method where the data set is reduced by redefining the feature
vectors and initialize the FCM algorithm with an enhanced cluster center formula.
The method aims at reducing the computational time of FCM. Laszlo Szilagyi [67]
developed a hybrid c-means algorithm as generalization of fuzzy, possibilistic, and
hard c-means algorithm. The hybrid model is induced with some novel image
segmentation method to form a virtual endoscopic procedure using MRI images of
brain. Shasidhar et al. [68] developed a modified FCM algorithm in segmentation
of brain MRI where a comprehensive feature vector space is used for the
segmentation technique. Comparative analysis in terms of segmentation efficiency
and convergence rate is performed between the conventional FCM and the
modified FCM. Saha et al. [69] describes a support vector machine (SVM) based
fuzzy clustering techniques and segments satellite images using the proposed
method. Mohapatra et al. [70] proposes a judicious integration of rough sets and
fuzzy sets suitably employed towards leukocyte segmentation in a clustering
framework. In their study, the goodness of fuzzy sets and rough sets is suitably
integrated to achieve improved segmentation performance. The membership
concept of fuzzy sets is efficient in handling of overlapping partitions, and the
rough sets provide a reasonable solution to deal with uncertainty, vagueness, and
�
�
���
�
incompleteness in data. Such synergistic combination gives the proposed scheme
an edge over standard cluster-based segmentation techniques, that is, K-means, K-
medoid, fuzzy c-means, and rough c-means. Wang et al. [71] put forwards a novel
two-dimensional fuzzy C-means (2DFCM) algorithm for the molecular image
segmentation. The 2DFCM algorithm is composed of three stages. The first stage
is the noise suppression by utilizing a method combining a Gaussian noise filter
and anisotropic diffusion techniques. The second stage is the texture energy
characterization using a Gabor wavelet method. The third stage is the feeding of
introduction of spatial constraints provided by the de-noised data and textural
information into the two-dimensional FCM clustering algorithm. The incorporation
of intensity and textural information allows the 2DFCM algorithm to produce
satisfactory segmentation results for images corrupted by noise and intensity
variations.
Beevi et al. [72] have developed a robust segmentation technique that
exploits histogram based FCM algorithm for the segmentation of medical images.
The algorithm first removes noise from the images and then segmentation is
performed. Sparse 3D transform-domain collaborative filtering is used to perform
denoising. Histogram is used to initialize the parameters of the FCM to avoid
convergence in local minima. Spatial probability is incorporated in the objective
function to boost the robustness of algorithm against noise. In the proposed
segmentation methodology, two types of spatial information are incorporated in the
membership function of FCM. First is Apriori probability and other is Fuzzy
Spatial Information. Apriori Probability is incorporated in the membership
function to assign a noisy pixel to a cluster that contains a large number of noisy
�
�
���
�
pixels in its neighborhood. Fuzzy Spatial Probability is incorporated in the
membership function and a pixel gets a higher membership value to a cluster when
its neighborhood pixels have high membership value to that cluster [72]. This
approach converges more quickly than the conventional FCM and attains reliable
segmentation accuracy apart from noise levels.
Krindis et al. [73] have developed Fuzzy Logic Information C-Means
Clustering (FLICM) algorithm. FLICM is introduced with a new factor in
objective function of FCM and the new factor ?@� has following characteristics
[73]:
• It incorporates local spatial and gray level information in a fuzzy way to
preserve robustness and noise insensitiveness.
• It controls the influence of the adjacent pixels depending on their distance
from middle pixel.
• It uses original image as input and avoids preprocessing steps to preserve
image details.
• It is free from any parameter selection.
This new novel fuzzy factor is defined mathematically as [73]
?@� � A �)�� B ��CD�
�E�
F� G H@�I6J�� G K@J6
Where ith pixel is the center of the local window e.g., 3*3, k is reference
cluster and jth pixel belongs to the set of neighbors that fall into a window around
ith pixel ���� and )�� is Euclidean distance between pixels i and j . The factor ?L�
�
�
���
�
is free from any control parameter for balancing image detail and noise. The
stability between noise and image details is automatically attained by the fuzziness
of every pixel of the image. The factor ?L� formulates the influence of the pixels
within the local window, to change flexibly according to their distance from the
middle pixel by using Euclidean distance that is�)��. It is important to note that the
factor ?L� reflects the damping extent of the neighbors with the spatial distances
from the central pixel. The modified objective function is defined as [73]
�M �AANO@�MP
@��Q�� G K@Q6 B ?@�
D
���
FLICM has some advantages over the FCM algorithm such as
insensitiveness towards noise and free from any parameter initialization.
Kannan et al. [74] have proposed a Novel Fuzzy Clustering C-Means
Algorithm (NFCM) for intensity inhomogeneities or weighted bias estimation and
segmentation of T1-T2 Brain MRI images of same patient. Authors also have
presented a center knowledge method to reduce the running time of the algorithm.
The Center Knowledge Algorithm rearranges the data matrix with respect to its
relabeling mean value and data is partitioned into c groups. A distance table is than
created to show the distance between the elements within each group. Maximum
distance between the groups is computed and a mean value is calculated. The
objective function has been modified for the said problem in [74] as follows
�M �AAH�@MNR)�@6 B S�)�@= �6T�
@��
P
���B �* N� GA)�@M
P
���T
�
�
���
�
where���)�@= � Q�@ B ; G U@ G K�Q6, ;�V��1��� and R� S 0 1
Objective function is minimized by employing same approach as that of
conventional FCM. �M is differentiated with respect to H�@, K� and U@ and is than
set equal to zero in order to get estimators for U, V and b. These estimators are then
used to design an algorithm to calculate tissue class and bias class. Lagrange
multiplier is then used to perform the task of membership evaluation. The
advantage of NFCM is that it can be applied at an early phase of automated data
analysis. NFCM is found to deal effectively with image intensity inhomogeneities
and noise present in the image.
Fuzzy clustering, over the years, has proved to be a powerful tool for image
segmentation. Most of the imaging techniques used in practice such satellite
imaging or medical imaging contain fair amount of uncertainty or vagueness.
Segmentation methods driven by crisp decision cannot handle this uncertainty
prudently. The subtle variations present in complex images are not too well
adapted to hard decisions. Fuzzy clustering algorithms can be extended and
modified according to the needs of a particular application. In medical imaging
various kinds of images exist, from radiological images to microscopic cytological
images. The bones and the tissue in radiological images or the cell and nucleus in
cytological images are hard to distinguish through hard segmentation. The
variations in intensity or gray level are so subtle that hard segmentation tends to
produce faulty result. The criterion or objective function and the flexible
membership function of fuzzy clustering efficiently deal with the situation. The
�
�
���
�
application of Fuzzy c-means (FCM) algorithm in cytological image segmentation
is studied and discussed in details in the coming chapters.