chapter iv comparative study on fuzzy membership...
TRANSCRIPT
53
CHAPTER IV
COMPARATIVE STUDY ON FUZZY MEMBERSHIP FUNCTIONS
4.1 Introduction
Objective of this chapter is identification of a fitting fuzzy membership degree
function. It is an essential task to define a suitable function to measure the global or local
image fuzzyness or ultrafuzzyness. In this backdrop a comparative study on fuzzy
membership functions for image segmentation using ultrafuzziness is carried out. In this
work, Tizhoosh membership function which is totally supervised, Huang & Wang
membership function and S-function are considered. Each membership function has its
own advantages and challenges in the computation process. Using fuzzy logic concepts,
the problems involved in finding the minimum or maximum of an entropy criterion
function are avoided. An attempt is made to identify the better membership function
through experiments to assign the fuzzy membership grade to every pixel in the image,
for optimum image segmentation using ultrafuzziness. For low contrast images contrast
enhancement is assumed. Experimental results demonstrate that there is a quantitative
improvement with S-function over the other two functions.
In many computer vision and image processing applications the fundamental task
performed on image data is image segmentation, which is done in order to process the
54
foreground objects and to explore the features. A fast and facile technique for image
segmentation is thresholding. The accuracy of segmentation depends upon the process
which is based on the gray level histogram. Fuzzy set theory provides better convergence
when compared with non-fuzzy methods. This thesis focuses mainly on an automated
method with fuzzy S-function and image ultrafuzziness as a fuzzy measure not
considering an entropic criterion function. In this scenario choosing better fuzzy
membership function is an essential task.
Fuzzy based Image Thresholding methods are introduced in literature to
overcome the grayness ambiguity problem. Fuzzy set theory is used in these methods to
handle grayness ambiguity or image vagueness during the process of threshold selection.
The gray level intensity value is selected as the optimum threshold at which the
ultrafuzzy index is minimized. The image is considered as a fuzzy set and the
membership distribution explains each pixel belongs to either objet set or background set
in the misclassification region of the histogram. Hamid R. Tizhoosh, introduced a new
fuzzy measure called ultrafuzziness and also a new a membership function.
4.2 Fuzzy Membership Functions
In this work, we considered three different fuzzy membership functions (MF)
such as:
Tizhoosh membership function.
S-function for membership grades.
Huang and Wang membership function.
55
These are some of the important fuzzy membership functions. In the following
sections each one is explained briefly and results are compared in chapter VI.
4.2.1 Fuzzy membership function of Tizhoosh
To measure the image fuzziness Hamid R. Tizhoosh [26] defined a new
membership degree function as shown in Equation (1) which comprises of three
unknown quantities α ,β and T must be estimated from the image statistics.
μ(g) = (4.1)
In this e experiment we have considered α ,β both values are equal to 2.
Fuzzy sets of type I
The most common measure of fuzziness is the linear index of fuzziness. For an
M N image subset A ⊆ X with gray levels g ⊆ [0, L-1], the linear index of fuzziness
can be estimated as follows:
(A) = (g), 1- (g)] (4.2)
Where ( g) is obtained from Equation(1). So the optimal threshold can be obtained
though maximizing the linear index of fuzziness criterion function that is given by
t*= Arg min { (A: T )}, 0 ≤ T ≤ L-1 (4.3)
56
Fuzzy sets of Type II
Definition: A type II fuzzy set à is defined by type II membership function X
where x∊ X and u ∊ ⊆ [0, 1]
à can be expressed in the notation of fuzzy set as
à = {(( , u), ( , u))| ∊ X, u ∊ ⊆ [0, 1]},
in which 0≤ (x,u) ≤ 1.
Figure 4.1 A possible way to construct type II fuzzy sets. The interval between lower/left
and upper/right membership values (bounded region) will capture the footprint of
uncertainty.
A type II fuzzy set can be defined from type I fuzzy set and assign upper and lower
membership degrees to each element to construct the footprint of uncertainty as shown in
Figure 1. A more suitable definition for a type II fuzzy set can be given as follow:
(4.4)
The upper and lower membership degrees and of initial membership
function µ can be defined by means of linguistic hedges like dilation and concentration:
57
,
,
Hence, the upper and lower membership values can be defined as follows:
,
,
Where δ ∊ (1, ∞) but δ > 2 is usually not meaningful for image data.
Ultrafuzziness
If the degrees of membership can be defined without any uncertainty as type I
fuzzy sets, then the ultrafuzziness should be zero. When individual membership values
can be indicated as an interval, the amount of ultrafuzziness would increase. The
maximum ultrafuzziness is one when the information of membership degree values are
totally ignored. For a type II fuzzy set, the ultrafuzziness is defined as γ for an M N
image subset à ⊆ X with gray levels g ⊆ [0, L-1], histogram h(g) and membership
function can be defined as follow:
(Ã)= (4.5)
where
,
, δ ∊ (1, 2)
58
4.2.2 S-Function
The existing method has several drawbacks in constructing the fuzzy membership
degree function. The three unknown quantities α, β and T are to be estimated from the
image statistical parameters of the image histogram. Since they vary from one image to
another it becomes difficult to automate the entire process of image thresholding. We
considered the standard S-function to compute membership degree of the fuzziness of the
given image. We derived a most convincing method to compute initial fuzzy seed subsets
as the S-function described in Figure 4.2.
Figure 4.2 Shape of the S-function.
The S- function is used for modeling the membership degrees for object pixels.
(x) = S (x; a, b, c) = (4.6)
59
Figure 4.3 Multimodal image histogram and the characteristic functions for the seed
subsets.
Iinitial fuzzy seed subset values a, b and c are computed based on. . Let x(i, j)
be the gray level intensity of image at (i, j). I={ x(i, j)| i∊ [1, M], j ∊ [1, N]} is an image
of size M N, i.e., n. The gray level set {0,1,2,…..255}. The mean(µ) and standard
deviation( ) are calculated as follows
µ = (4.7)
= (4.8)
From Equations (4.7) and (4.8) fuzzy seed set values a, b and c as shown in figure 4.3,
are estimated as follows:
b = µ = (4.9 )
a = µ - (4.10)
c = µ + (4.11)
Kaufmann introduced an index of fuzziness first to measure the vagueness of a fuzzy set.
He also established four conditions that every measure of fuzziness should satisfy. This
60
fuzzy measure, ultrafuzziness is also satisfying his four conditions. So the optimal
threshold can be obtained though minimizing the ultrafuzziness criterion function that is
given by
t*= Arg min { (Ã: T )}, 0 ≤ T ≤ L-1 (4.12)
4.2.3 Fuzzy membership Function of Huang and Wang
Let X denotes an image set of size M x N with L levels, and is the gray level
of a (m, n) pixel in X. Let denote the membership value which represents the
degree of possessing a certain property by the (m, n) pixel in X; that is, a fuzzy subset of
the image set X is a mapping μ from X into the interval[0,1]. In this notation of fuzzy set,
the image set X can be written as
X= {( ( ))}, ( ) where 0 ≤ ( ≤ 1, m=0, 1,… M-1 and n=0, 1,…,N-1.
Here, the membership function can be viewed as a characteristic function that
represents the fuzziness of a (m, n) pixel in X. For the purpose of image thresholding,
each pixel in the image should possess a relationship between the pixel and its belonging
region.
Let h(g) denote the number of occurrences at the gray level g in an input image. Given a
certain threshold value t, the average gray levels of the background and the object
can be, respectively, obtained as follows:
= / (4.13)
= / (4.14)
61
The average gray levels, and , can be considered as the target values of the
background and the object for the given threshold value t. The relationship between a
pixel in X and its belonging region should intuitively depend on difference of its gray
level and the target value of its belonging region. Thus, let the relationship possess the
property that the smaller the absolute difference between the gray value of a pixel and its
target value is, the larger the membership value that pixel has. Hence, the membership
function which evaluates the above relationship for a (m, n) pixel can be defined as:
( ) = (4.15)
Where C0 and C1 are two constants such that 0 ≤ ( ≤ 1. For the given threshold t,
any pixel in the input image should belong to either the object or the background. Hence
it is expected that the membership value of any pixel should be no less than ½. The
membership function in ( ) really reflects the relationship of a pixel with its belonging
region.
4.3 General Thresholding algorithm for type II fuzzy sets with
ultrafuzziness
The general algorithm for image thresholding based on type II fuzzy sets and a
fuzzy measure called Ultrafuzziness can be formulated as follows:
1 Select
any one membership function.
62
2 Comput
e membership degree for every pixel.
3 Calculat
e image histogram.
4 Initializ
e the position of the membership function.
5 Shift
the membership function along the gray-level range.
6 Calculat
e in each position the amount of ultrafuzziness from Equation(5).
7 Find
out the position with minimum ultrafuzziness.
8 Thresho
ld the image with t* = .
4.4 Experimental trials
The comparison work involved with lot of experimentation by considering a
suitable data set which comprises of various kinds of images. Actually, effectiveness and
challenges of each function is figured out in the upcoming chapter called results and
discussions. Here, it is tried to show some images and their thresholded images as a result
of using different membership functions.
63
(a)
(b) (c)
(d) (e)
Figure 4.4 (a) Original image of Bacteria. (b) Its ground truth image. (c) Thresholded
image using Tzhoosh function. (d) Thresholded image using Wang function. (e)
Thresholded image using S- function.
64
(a)
(b) (c)
(d) (e)
Figure 4.5 (a) Original image of Rhino. (b) Its ground truth image. (c) Thresholded image
using Tizhoosh function. (d) Thresholded image using Wang function. (e) Thresholded
image using S-function.
65
(a)
(b) (c)
(d) (e)
Figure 4.6 (a) Original Ultrasound image (b) Its ground truth image (c) Thresholded
image using Tizhoosh function. (d) Thresholded image using Wang function. (e)
Thresholded image using S-function.
66
4.5 Chapter summary
This chapter provides an overview about fuzzy membership functions and their
significance in fuzzy oriented image segmentation approaches. In this perspective we
considered S-function along with other membership functions. Tizhoosh introduced a
new fuzzy membership function with many parameters and needs a great supervision for
every image considered. Therefore the process is truly manual and hard to get the
threshold. Wang membership function may be simple compared to Tizhoosh function,
however, it involves more complexity. S-function is more simplistic and demands less
computational complexity with our simplification of linguistic hedges based on image
basic characteristics. But this method can be further improved and tested against other
fuzzy membership functions available or a still new suitable membership function can be
derived. Efficiency of threshold selection is demonstrated with experimental results. We
assume a reasonable contrast enhancement for low contrast images. Performance
evolution is carried out with the help of two popular approaches, misclassification error
and Jaccard Index on the proposed work in the chapter titled „results and discussions‟.
Owing to the computational limitations all the other fuzzy membership functions
available in the literature are not tested. The other available functions can also be tested
against popular S-function. With this work we conclude that the performance of S-
function is relatively good over the other two membership functions and so here
employed this function.