You are on page 1of 24

42

Chapter 2
Fundamental tools for image retrieval
2.1. Introduction
The principal task of image retrieval is to measure the relevance between shapes,
which are represented by their features. Therefore, three steps are essential in shape based
image retrieval. First, the extraction of features using some shape descriptor. Second, the
measuring of relevance between the extracted features of images through some similarity
measure. Third, their performance evaluation using some performance measure. Thus, in
this chapter, we describe fundamental tools for feature extraction, similarity measures, and
performance measures. Since the core theme of our research is based on radial moments,
for feature extraction, we use radial moments, which are described along with their
properties in the first part of this chapter. Afterwards, we describe various similarity
measures, which are used to measure similarity among images. Besides, we also describe
performance measurements for evaluating the performance of various methods for image
retrieval. The contributions of this chapter include the following:
- To describe various radial moments along with their properties.
- To describe various image databases used for the evaluation of image retrieval
system.
- To describe various similarity measures and performance measures for the
evaluation of image retrieval system.
- To evaluate the performance of radial moments in terms of image retrieval.
- To analyze the performance of various similarity measures.

2.2. Feature extraction tools
Moments are scalar quantities used to characterize a function and to capture its
significant features. Moments have been used to differentiate shapes of various characters,
aircrafts, chromosomes, etc. Moreover, these types of image features provide a complete
object representation that is invariant to similarity transformations. In mathematical form,
moments are the projection of a function on to polynomials basis. In the following sub-
sections various radial moments are described and discussed.

43

2.2.1. Geometric moments (GMs)
Simple geometric properties of an image, such as area, position, and orientation
can be easily computed from a set of linear functions of the image called geometric
moments. Geometric moments are also called regular moments. The geometric moments
of order ) ( q p + of image intensity function ) , ( y x f are defined as

, ) , ( dxdy y x f y
x m
q
p
pq } }
=


(2.1)

where . ..., , 2 , 1 , 0 , = q p The above definition has the form of projection of the function
) , ( y x f onto the monomials y
x
q
p
. However, the basis set } { y
x
q
p
is not orthogonal. The
moments upto order two provide basic geometric characteristics of ) , ( y x f . The moment
m00
defines the total mass of ) , ( y x f , whereas the moments ) , , (
01 10
00
m m
m , are sufficient
to compute the centre of mass, i.e., centroid ) , ( y
x
c
c
of the image ) , ( y x f , given as:

, ,
00
01
0 0
10
m
m
y
m
m
x
c
c
= = (2.2)

A central moment corresponding to
mpq
is defined as:

, ) , ( ) ( ) ( dxdy y x f y y
x
x
c
q
c
p
pq

} }
=


(2.3)

The central moments are equivalent to the regular moments of an image that has been
shifted such that the image centroid ) , ( y
x
c
c
is at the origin. As a result, central moments
are invariant to translation of the image.

2.2.2. Moments invariants (MI)
The use of moments for image analysis and object recognition was initiated by Hu
[47] and described in detail in [69-71]. Using the linear combination of geometric
moments, Hu derived a set of invariant moments, which possess the desirable properties of
44

being invariant under translation, scaling, and rotation. From the second and third order
geometric moments, we have the following seven orthogonal invariants:

], ) ( ) ( 3 [ ) ( )
3
(
] ) ( 3 ) ( [ ) )(
3
(
) )( ( 4 ] ) ( ) ( )[ (
] ) ( ) ( 3 [ ) ( )
3
(
] ) ( 3 ) ( [ ) ( )
3
(
) ( ) (
)
3
( ) 3 (
4 ) (
03 21
2
12 30
2
03 21 30 12
03 21
2
12 30
2
12 30 03 21
7
03 21 12 30 11 03 21
2
12 30
2
02 20
6
03 21
2
12 30
2
03 21 03 21
03 21
2
12 30
2
12 30 12 30 5
03 21
2
12 30
2
4
03 21
2
12 30
2
3
2
11 02 20
2
2
02 20
1
m m m m m m m m
m m m m m m m m
m m m m m m m m m m m
m m m m m m m m
m m m m m m m m
m m m m
m m m m
m m m
m m
+
+
+ +
+ + + =
+ + + =
+ + + +
+ + + + =
+ =
+ =
+ =
+ =
|
|
|
|
|
|
|
(2.4)

where |
7
is the skew orthogonal moment. This set of moments is invariant to translation,
scale, mirroring (with a negative sign) and rotation.

2.2.3. Legendre moments (LMs)
The Legendre moments of order ) ( q p +

for an image function ) , ( y x f are defined
as:

, ) , ( ) ( ) (
4
) 1 2 )( 1 2 (
} }
+ +
=


dxdy y x f y
P
x
P
q p
q p pq
(2.5)

where . ..., , 2 , 1 , 0 , = q p The Legendre polynomials )} ( { x
Pp
[72] are a complete
orthogonal basis set over the interval ] 1 , 1 [ .

,
1 2
2
) ( ) (
1
1
o pq q p
p
x
P
x
P
+
=
}

(2.6)

where p
th
order Legendre polynomial is given by [73]:
) 1 (
!
2
1
) (
2
=
x
dx
d
P
x
P
p
p
p
p
p
(2.7)
45

2.2.4. Rotational moments (RMs)
The rotational moments of order ) ( q p +

for an image function ) , ( y x f are defined
as:
}}
=
s +

1
2 2
, ) , (
y x
jq p
pq
dxdy y x f
e r D
u
(2.8)

where y
x
r
2
2
+ = , = ..., , 2 , 1 , 0 p and q takes any positive and negative integer values
and ( ).
tan
1
x y

= u

2.2.5. Complex moments (CMs)

The notion of complex moments was introduced as a simple and straightforward
way to describe rotation invariant moments. The complex moments of order ) ( q p +

for an
image function ) , ( y x f

are defined as:

}}
=
s +
+
1
2 2
) (
, ) , (
y x
q p j q p
pq
dxdy y x f
e r C
u
(2.9)

where y
x
r
2
2
+ = , = ..., , 2 , 1 , 0 , q p

and 1 = j and ( ).
tan
1
x y

= u

2.2.6. Zernike moments (ZMs)
Teague [49] has suggested the use of orthogonal moments in image analysis based
on orthogonal polynomials. This set of orthogonal functions has been introduced by
Teague as a basic tool for representation of a wavefront function for optical systems with
circular pupils. Since then radial polynomials have been found important in applications
ranging from pattern recognition, optical engineering, medical imaging to eye diagnostics,
etc [74-79]. The radial orthogonal moments are defined on a unit circle, i.e., the image
plane is } 1 ) , {(
2
2
s + = = y
x
y x D . The set of orthogonal ZMs for an image intensity
function ) , ( y x f with order p

and repetition q are defined over a continuous unit disc
[49]:

46

}}
+
=
s + 1
2 2
*
, ) , ( ) , (
1
y x
pq pq
dxdy y x
V
y x f
p
Z
t
(2.10)

where ) , (
*
y x
V pq
is the complex conjugate of the Zernike polynomials ) , ( y x
V pq
, defined
as:

, ) ( ) , (
e
r
R
y x
V
jq
pq pq
u
= (2.11)

where ( ).
tan
, 1 , , 0 , 0
1
x y and j even q p p q p

= = = s s > u The radial polynomials
) (r
Rpq
are defined by:

( )
,
!
2
!
2
!
! ) (
) 1 ( ) (
2
2
0
r
k
q p
k
q p
k
k p
r
R
k p
q p
k
k
pq

=
|
|
.
|

\
|

|
|
.
|

\
|

+

= (2.12)

The radial polynomials ) (r
Rpq
satisfy the orthogonality relation:

,
) 1 ( 2
1
) ( ) (
' '
1
0
o pp q p pq
p
dr r r
R
r
R
+
=
}
(2.13)

where
o ij
is Kronecker delta. The set of Zernike polynomials ) , ( y x
V pq
forms a complete
orthogonal set within the unit disc as:

.
1
) , ( ) , (
' '
*
' '
2
0
1
0
o o
t
u
t
qq pp q p pq
p
rdrd y x
V
y x
V
+
=
} }
(2.14)

Features based on orthogonal kernels are more effective in terms of information
compactness and minimal information redundancy. Therefore, the number of moments
required to reconstruct an image is much less than those obtained from non-orthogonal
kernels. The image function is reconstructed by using inverse moments and transforms,
which is given as:

47

). , ( ) , (

max
0
y x
V Z
y x f
pq
p
p
p
p q
pq
=
= =
(2.15)

2.2.7. Pseudo Zernike moments (PZMs)
PZMs also belong to the class of circularly orthogonal moments, which are also
defined over a unit disc [50]. These moments have similar characteristics as that of ZMs
and are widely used and studied in the literature because of their minimum information
redundancy and immunity to noise. The computational cost of PZMs is more than ZMs for
the same order, which is also demonstrated in the later section of this chapter. On the other
hand, PZMs are observed to be more robust to image noise than ZMs [51]. In addition,
PZMs provide twice the number of moments as provided by ZMs for the same moment
order. There are ) 1 (
max
2
p +

number of PZMs as compared to 2 ) 2 )( 1 (
max max
p p + +
number of ZMs for the same maximum order p
max
. Thus, using the same maximum order
p
max
, PZMs have more low-order moments than ZMs. As a result, PZMs are less sensitive
to image noise than ZMs. PZMs differ from ZMs in their real valued radial polynomials
defined as [51]:

( )
( ) ( )
,
! ! 1 !
! 1 2 ) 1 (
) (
0

+ +
+
=

=

q p
s
s p
s
pq
s q p s q p s
r
s p
r
R
(2.16)

where p q p s s > 0 , 0 . The orthogonality principles for PZMs are similar to that of ZMs
given by Eq. (2.13) and Eq. (2.14). The image reconstruction function for PZMs is similar
to that of ZMs given by Eq. (2.15) with the replacement of
Z pq
by
PZ pq
, where

}}
+
=
s + 1
2 2
*
) , ( ) , (
1
y x
pq pq
dxdy y x
V
y x f
p
PZ
t
(2.17)

2.2.8. Orthogonal Fourier Mellin moments (OFMMs)
The new orthogonal radial polynomials have more zeros than do the Zernike radial
polynomials in the region of a small radial distance [52]. OFMMs may be thought of as
generalized Zernike moments and orthogonalized complex moments. For small images,
48

the description by OFMMs is better than that of ZMs in terms of image-reconstruction
errors and signal-to-noise ratio. The OFMMs basis functions also form a set of complete
orthogonal functions over the unit disc and they differ from ZMs in their polynomials
defined as:

}}
+
=
s + 1
2 2
*
, ) , ( ) , (
1
y x
pq pq
dxdy y x
V
y x f
p
O
t
(2.18)

where

, ) ( ) , (
e
r
R
y x
V
jq
p pq
u
=
(2.19)

with constraints 0 , 0 > > q p . It is observed that p does not depend q on unlike ZMs and
PZMs. The orthogonal radial polynomials are defined as:

,
! ) 1 ( ! ! ) (
! ) 1 (
) 1 ( ) (
0
r
s s s p
s p
r
R
s
p
s
s p
p

+
+ +
=
=
+
(2.20)

) (r
Rp

is orthogonal over the interval 1 0 s sr , i.e.,

( )
.
1 2
1
) ( ) (
1
0
o pq q p
p
dr r r
R
r
R
+
=
}
(2.21)

Orthogonality principles for OFMMs polynomials are similar to that of ZMs given by
Eq. (2.14). The image can be reconstructed by using the following expression
) , ( ) , (

max
0
max
max
y x
V O
y x f
pq
p
p
q
q
pq
q
=
= =
(2.22)
where p and q are the maximum order and repetition.



49

2.2.9. Radial harmonic Fourier moments (RHFMs)
RHFMs basis functions also form a set of complete orthogonal functions defined
over a unit disc. RHFMs of order p and repetition q with 0 > p and 0 > q are defined as
[56]:
}}
=
O
s + 1
2 2
*
, ) , ( ) , (
2
1
y x
pq pq
dxdy y x
V
y x f
t
(2.23)
where
, ) ( ) , (
e
r
R
y x
V
jq
p pq
u
= (2.24)

The radial kernel function is defined as:

+
=
=
odd p r p
r
even p pr
r
p
r
r
Rp
, ) ) 1 ( sin(
2
, ) cos(
2
0 ,
1
) (
t
t (2.25)
The orthogonal property for radial kernel is given as:

. ) ( ) (
1
0
o pq q p
rdr r
R
r
R
=
}

(2.26)

The orthogonality of basis function is given as:

. 2 ) , ( ) , (
' '
*
' '
2
0
1
0
o o
t u
t
qq pp q p pq
rdrd y x
V
y x
V
=
} }
(2.27)

The reconstruction function for RHFMs is similar to that of OFMMs given by Eq. (2.22),
with the replacement of
Opq
by
Opq
.



50

2.2.10. Chebyshev Fourier moments (CHFMs)
CHFMs of order p and repetition q with 0 0 > > q and p

are defined as [57]:
}}
=
+
s + 1
2 2
*
, ) , ( ) , (
2
1
y x
pq pq
dxdy y x
V
y x f
t

(2.28)

where ) , (
*
y x
V pq
is the complex conjugate of the basis function ) , ( y x
V pq
defined by:
, ) ( ) , (
e
r
R
y x
V
jq
p pq
u
= (2.29)

where the radial kernel function is given as:


(2.30)

The orthogonal properties for radial kernel and basis function are similar to that of
RHFMs given by Eq. (2.26) and Eq. (2.27), respectively. The reconstruction function for
CHFMs is similar to that of OFMMs as given by Eq. (2.22), with the replacement of
Opq
by
+pq
.

2.3. Properties of radial moments

2.3.1. Orthogonal property
From various described moments in the previous section GMs, MI, RMs, and CMs
are not orthogonal and rest of the described moments satisfy orthogonality property. By
orthogonality, we mean the decomposition of an object into uncorrelated components to
simplify its analysis. The lack of orthogonality in GMs corresponds to high correlation
among moments, which leads to more redundant information of the image. Besides, non
orthogonality implies inadequacy of information compactness in each of the computed
moments. On the other hand, orthogonality of the kernels means that an image is projected
onto a set of pairwise orthogonal axes, and the classifier can hence be relatively simple

( ) ) 1 2 ( 2
! ) 2 ( !
! ) (
) 1 (
1 8
) (
2
2 /
0
4 / 1

|
.
|

\
|
=

=
r
k p k
k p
r
r
r
R
k p
p
k
k
p
t
51

[51]. The orthogonality principles satisfy by LMs, ZMs, PZMs, OFMMs, RHFMs, and
CHFMs are given in the previous section.

2.3.2. Invariance property
The set of complex radial moments inherently possess rotation invariance property.
The magnitude values of ZMs, PZMs, OFMMs, RHFMs, CHFMs remain identical before
and after rotation. Therefore, the magnitude values of above mentioned moments are
rotation invariant. Here, we describe the rotation invariance property in terms of ZMs,
which is similar for all other radial moments. For an image rotated by an angle o , ZMs
are defined as
,
e Z Z
jq
pq
r
pq
o
= (2.31)
where
Z
r
pq
and
Z pq
are ZMs of rotated and unrotated image, respectively. The rotation
invariant ZMs are extracted by considering only magnitude values as

e Z Z
jq
pq
r
pq
o
=
(2.32)


1 ) sin( ) ( cos = + =

o o
o
q j q
e
jq

(2.33)

Substituting Eq. (2.33) in Eq. (2.32), we get

Z Z pq
r
pq
=
(2.34)
As Z
Z pq
q p
=

*
,
and
Z Z q p pq
=
,
, therefore only magnitudes of ZMs with 0 > q are
considered [95]. However, ZMs are not scale and translation invariant. Scaling and
translation invariance can be achieved by normalizing the image, which we discuss in the
later Chapter. LMs possess the orthogonality property. However, they do not exhibit
rotation invariance.

2.3.3. Robustness to noise
The computation of orthogonal radial moments is based on the summation
operation. We know that an image processing operation based on summation or
52

integration is robust to noise. However, the robustness of various radial moments varies
from each other. PZMs and OFMMs exhibit more feature vectors than ZMs for the same
order of moment. Therefore, PZMs are less sensitive to image noise as compared to ZMs.
In our experiments, we evaluate the performance of all radial moments for robustness to
noise, which is given in Section 2.8.


(a)


Moments
10 20 30 40
ZMs

PZMs

OFMMs

RHFMs

CHFMs

(b)
Fig. 2.1 (a) Original image (b) reconstructed images by moments at orders
40 , 30 , 20 , 10
max
= p


2.3.4. Multilayer expression
Low orders of orthogonal radial moments represent the global information of the
image and high orders of moments represent the details of the image. A binary image of
the character A of size 64 64 pixels, given in Fig. 2.1 (a) is reconstructed for
p
max
= 10, 20, 30 and 40 for ZMs, PZMs, OFMMs, RHFMs, and CHFMs. The
reconstruction capability of various moments is measured by visual inspection of the
p
max
=q
max
53

reconstructed images. Fig. 2.1(b) displays the reconstructed images of the original image
of character A, which is given in Fig. 2.1(a). Besides, the total number of moments L
required to reconstruct an image for each moment are presented in Table 1. It is seen from
Fig. 2.1(b) that the low orders of moments provide the gross approximation of image and
there is a gradual improvement in the quality of reconstructed image as the moment order
increases. We also observe that PZMs and OFMMs become numerical unstable for
moment orders 20
max
> p . Consequently, their image reconstruction capability diminishes
for moment orders 20
max
> p .

Moments L 10 20 30 40
ZMs
) 2 )( 1 ( 2 1
max max
p p + +

66 231 496 861
PZMs
) 1 (
max
2
p +

121 441 961 1681
OFMMs ) 2 1 )( 1 (
max max
q p + +

231 861 1891 3321
RHFMs ) 2 1 )( 1 (
max max
q p + +

231 861 1891 3321
CHFMs ) 2 1 )( 1 (
max max
q p + +

231 861 1891 3321
Table 2.1 Total numbers of moments L , for 40 , 30 , 20 , 10
max max
= = q p



Fig. 2.2 Comparison of CPU elapsed time for the computation of ZMs, PZMs, OFMMs,
RHFMs, and CHFMs by applying recursive algorithms

2.3.5. Derivation of higher order moments
The higher order orthogonal radial moments, such as ZMs, PZMs, OFMMs,
RHFMs, and CHFMs, are easy to derive unlike Hus moment invariants. Hus moment
invariants are derived from three low orders of GMs. However, the question of what is
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
2 4 6 8 10 12
C
P
U

t
i
m
e

(
s
e
c
)
p
max
ZM
PZM
OFMM
RHFM
CHFM
54

gained by including higher order moments in the context of image analysis has not been
addressed by Hu. Besides, the recovery of image form these moments is supposed to be a
difficult task. As we can see in Fig. 2.1 that orthogonal radial moments have the strong
capability to reconstruct an image, which demonstrates the image representation and
description capability of the features extracted by these moments. Nevertheless, the radial
moments are computation intensive due to the inclusion of factorial terms and polynomials
of order p that contains q number of terms. To address this issue, various recursive
algorithms [80-85] are available, which considerably enhance the speed of computation.
Besides, the recursive methods for the computation of radial moments are numerically
stable for very high orders of moments, e.g., the recursive q

method [81] for
computation of ZMs is numerically stable even for moment order 300
max
= p for an image
of size 256 256 pixels. In our experiments, we implement ZMs, PZMs, OFMMs,
RHFMs, and CHFMs by using recursive q

method and compare their CPU elapsed
time. The CPU time comparison for the computation of ZMs, PZMs, OFMMs, RHFMs,
and CHFMs at various orders , 10 , 8 , 6 , 4 , 2
max
= p

and 12 is presented in Fig. 2.2. It is
observed from the figure that OFMMs, RHFMs, and CHFMs takes the highest and similar
CPU time for their computation, followed by PZMs and ZMs at the same moment order.
In other words, we can say that ZMs take least amount of CPU time.

2.4. Similarity measures

An effective similarity measure is required because it plays a significant role in
improving the retrieval accuracy of the system. The key task of similarity measure is to
classify the images based on their relevance to query image. The similarity measures are
also termed as distance metric or similarity metric. Similarity measures are used to
determine the relevance between the query image and images stored in the database. The
database image with the smallest distance to query image is referred to as the most
relevant image. Its inappropriate selection may lead to undesirable results. Several
similarity measures have been proposed by researchers to improve the image retrieval rate.
Various similarity measures are in existence such as Euclidean distance, Manhattan
distance, Bray-Curtis distance, Hellinger distance, Chebyshev distance, Canberra distance,
Chi-square statistics and cosine distance, where Euclidean distance is commonly applied
for similarity matching of images. In this sub-section, the objective is to analyze the
performance of above mentioned similarity measures on the image retrieval system, so
55

that some of the measures can be applied for an effective image retrieval task. Besides, we
also suggest a new similarity measure, which is partly inspired by histogram intersection.
Histogram intersection is basically applied for color histograms in order to match color
images as given by [86]. However, in our approach, we apply the modified form of
histogram intersection on the spectral features and name it as a min-max similarity
measure. Let a feature set contains F features, and ) (Q
Mi
and ) (D
Mi
represent the i
th

feature of query image Q and database image D, respectively. The various similarity
measures are described as:
2.4.1. Euclidean distance
This distance is one of the forms of Minkowski distance, which is based on
Lp

norm. For p=2 it is referred to as L
2
norm and also known as Euclidean distance. This
measure works in vector space to compute distance between two vector inputs. It is a
consequence of Pythagorean Theorem [87]. Euclidean distance is the most commonly
used measure. Mathematically, it can be expressed as:

( ) =

=
1
0
2
) ( ) ( ) , (
F
i
i i
D
M
Q
M
D Q d (2.35)
2.4.2. Manhattan distance
This measure is another form of Minkowski distance when p=1 and therefore,
known as L
1
norm. It represents distance between points in a city road grid. It determines
absolute distances between coordinates of pairs of vectors [88]. It is also known as city
block distance. Mathematically, Manhattan distance is defined as:

=

=
1
0
) ( ) ( ) , (
F
i
i i
D
M
Q
M
D Q d

(2.36)
2.4.3 Bray-Curtis distance
Bray-Curtis distance examines space as a grid similar to Manhattan distance with
the property that if all coordinates are positive, its value lies between 0 and 1. Zero Bray-
Curtis distance represents exact similar coordinate. The normalization is done using the
56

absolute difference divided by the summation [89,90]. It is also known as Sorensens
distance. Bray-Curtis distance can be defined as:

+

=

=

=
1
0
1
0
) ( ) (
) ( ) (
) , (
F
i
i i
F
i
i i
D
M
Q
M
D
M
Q
M
D Q d

(2.37)
2.4.4. Hellinger distance
This measure quantifies the deviation between two probability measures. It is
computational intensive. It has natural upper and lower bounds, 0 and 1, respectively. It is
based on the proportion of the protocol attributes [91]. It can be formulated as:

( )
=

=
1
0
2
2
) ( ) (
2
1
) , (
F
i
i i
D
M
Q
M
D Q
d

(2.38)

2.4.5. Chebyshev distance
It is also called as maximum value or chessboard distance. It examines the absolute
magnitude of the differences between coordinates of a pair of objects. This distance can be
used for quantitative and ordinal variables [92]. Mathematically, it can be defined as:

. 1 ..., , 1 , 0 , ) ( ) ( max ) , ( = = F i D
M
Q M D Q d
i
i

(2.39)

2.4.6. Canberra distance
Canberra distance examines the sum of series of fraction differences between
coordinates of pairs of objects. Each term of fractional difference has a value between 0
and 1. If one of the coordinate is zero, the term becomes unity regardless of other value,
thus distance will not be affected. Note that if both distances are zeros, we need to define
0/0=0. This distance is quite sensitive to small change when both coordinates are near to
zero [93]. It is defined as:

=

=
1
0 ) ( ) (
) ( ) (
) , (
F
i
i i
i i
D
M
Q
M
D
M
Q
M
D Q d (2.40)
57

2.4.7. _
2
(Chi-square) statistics
The _
2
distance is used to analyze the observed difference of frequencies from the
expected frequencies [90]. The _
2
distance is quasi distance and is defined as:

( )


=

=
1
0
2
) (
) ( ) (
) , (
F
i
i
i i
Q
M
Q
M
D
M
D Q d

(2.41)
2.4.8. Cosine distance
The cosine distance measures the distance in angular direction [94]. The distance is
measured by the angle between two vectors. The higher the angular separation the more
will be the similarity. It is also called coefficient of correlation. It is defined as:

( )

=

=
1
0
2
1
0
2
1
0
) ( ) (
) ( ) (
) , (
F
i
i
F
i
i
F
i
i i
D
M
Q
M
D
M
Q
M
D Q d (2.42)

2.5. Proposed similarity measure
In this section, we provide a novel similarity measure which is simple to compute
and gives effective performance. It is based on histogram intersection [86]. The histogram
intersection is generally applied for measuring similarity among color histograms.
However, the modified form is applied for the spectral features. The modification is
performed in the normalization part. Mathematically, it is expressed as follows:

( )
( )
,
) ( , ) ( max
) ( , ) ( min
1 ) , (
1
0
1
0

=
=
F
i
i i
F
i
i i
D
M
Q
M
D
M
Q
M
D Q d (2.43)

where the minimum function value is used to compute the similarity/dissimilarity among
images. It computes the minimum possible distance between query and database image.
For a database image D which is dissimilar to the query image Q, the expression
( )

=
1
0
) ( , ) ( min
F
i
i i
D
M
Q
M
yields the smaller distance as compared to similar images. On the
58

other hand, the expression ( )

=
1
0
) ( , ) ( max
F
i
i i
D
M
Q
M
yields the maximum possible value,
which effectively helps to normalize the feature components and the measured distance
always lies in [0, 1]. Therefore, by subtracting the computed distance from 1 is capable of
differentiating intra-class and interclass images, which places large disparity between
dissimilar images. We refer to the proposed similarity measure as the min-max similarity
measure.
2.5.1. Effectiveness of the proposed similarity measure
We present the classification performance of min-max similarity measure by
considering three query images camel-3, device7-10, and apple-3 taken from
MPEG-7 database (described in Appendix A). Their results are provided in Fig. 2.3(a),
2.3(b), and 2.3(c). It is apparent from the Fig. 2.3(a) that all the samples of camel class are
classified separately and images from other classes deer, elephant, and dog, which are
visually nearby to query image of the class camel are misclassified and intermixed among


(a)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 2 4 6 8 10 12 14 16 18 20
M
i
n
-
M
a
x

d
i
s
t
a
n
c
e
Number of samples in a class
camel
deer
dog
elephant
59


(b)

(c)
Fig. 2.3 Classification performance of min-max similarity measure for query image (a)
camel-3 (b) device07-10 and (c) apple-11

each other. The similar trends are observed while examining the performance of Min-Max
for device07-10, and apple-3 query images, which are demonstrated in Fig. 2.3(b) and
2.3(c), respectively.

2.6. Performance measures

We use two performance evaluation tests to measure the retrieval accuracy of the
system. The first one is precision ( P ) and recall ( R ) to evaluate the retrieval effectiveness
of the methods. The second test is the Bulls eye performance ) (BEP , which measures the
retrieval accuracy of the methods. These performance measures are described as follows
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 2 4 6 8 10 12 14 16 18 20
M
i
n
-
M
a
x

d
i
s
t
a
n
c
e
Number of samples in a class
device-07
device-02
device-05
device-01
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 2 4 6 8 10 12 14 16 18 20
M
i
n
-
M
a
x

d
i
s
t
a
n
c
e
Number of samples in a class
apple
bat
pocket
device-09
60

2.6.1. Precision and recall ) ( R P
Precision measures the retrieval accuracy and recall measures the ability to retrieve
relevant images from the database. Precision and recall are inversely proportional to each
other as the precision reduces, the recall increases. We use average precision and average
recall for all the retrieval results. For a query image Q, we compute precision and recall in
percentage as follows:
, 100 , 100 = =
D
n
R
T
n
P
Q
Q
Q
Q
(2.44)

where
nQ
represents the number of similar images retrieved from the database,
TQ

represents total number of images retrieved,
DQ
represents number of images in database
similar to query image Q.
2.6.2. Bulls eye performance ) (BEP

BEP is measured by the correct retrievals among the top R 2 retrievals, where R is
the number of shapes which are relevant to the query image in the database. We use
average percentage value to measure BEP.


2.7. Database construction

The description of all the standard image databases used for evaluating the image
retrieval performance throughout the research work are described thoroughly in
Appendix A.
2.8. Performance evaluation of radial moments for image retrieval
In this section, we evaluate the performance of radial moments, viz., RMs, CMs,
ZMs, PZMs, OFMMs, RHFMs, and CHFMs for image retrieval. Experiments are
performed on an Intel Pentium Core 2 Duo 2.10 GHz processor with 3 GB RAM.
Implementation is performed on VC++ 9.0. We consider five databases, MPEG-7 CE
Shape 1 Part-B, Kimia-99, COIL-100, Noise, and Rotation, for assessing the performance
of above mentioned moments. The information regarding above mentioned databases can
be obtained in Appendix A. The R P curves are used to examine their functioning on
these databases, which represent various sorts of images. In the first test, we analyze their
61

behavior for MPEG-7 database and the corresponding curves are presented in Fig. 2.4(a).
We observe that among all the moments, the worst performance is given by CMs, i.e.,
complex moments and the best performance is attained by ZMs and PZMs and their R P

curves overlap with each other. Therefore, we see that ZMs and PZMs are capable of
identifying images from the database in which large variation exists within the instances of
a class.
The second test is performed for Kimia-99 database and the corresponding R P

curves are depicted in Fig. 2.4(b). We observe that CMs again gives the poorest
performance and the highest performance is given by ZMs and CHFMs and their R P

curves almost overlap with each other. PZMs and OFMMs perform almost similar and
their R P curves coincide. By observing the R P curves, we can say that ZMs are
again competent to retrieve similar images from the database, which contains partial
occluded and distorted shapes. The third test is performed for COIL-100 database and the
R P

curves for all employed moments are demonstrated in Fig. 2.4(c). It is observed that
the order of performance of all the moments is similar as that for MPEG-7 database. Again
ZMs and PZMs perform superior to rest of the moments and poorest performance is given
by CMs. The fourth test is performed for evaluating the robustness of moments against
noise. The results are given in Fig. 2.4(d). It represents that PZMs, RHFMs, and CHFMs
moments achieve the highest robustness against noise followed by ZMs, OFMMs, RMs,
and CMs. However, the performance of ZMs and OFMMs is far superior to that of RMs
and CMs. Besides, a minute discrepancy is incurred in the performance of PZMs and ZMs.
CMs again gives the worst performance for retrieving noise affected images. In the fifth
test, rotation invariance of all the moments is analyzed and the results are illustrated in
Fig. 2.4(e). It can be seen that all the moments are highly invariant to rotation and achieve
100% precision for all the moments. Consequently, the R P

curves of all the employed
moments overlap with each other.


62


(a) (b)

(c) (d)


(e)
Fig. 2.4 R P comparisons of radial moments for (a) MPEG-7 Shape 1 Part-B (b)
Kimia-99 (c) COIL-100 (d) Noise and (e) Rotation databases

10
20
30
40
50
60
70
80
90
100
10 20 30 40 50 60 70 80
P
r
e
c
i

s
i
o
n

(
%
)
Recall (%)
MPEG-7
RMs
CMs
ZMs
PZMs
OFMMs
RHFMs
CHFMs
10
20
30
40
50
60
70
80
90
100
10 20 30 40 50 60 70 80 90 100
P
r
e
c
i

s
i
o
n

(
%
)
Recall (%)
Kimia-99
RMs
CMs
ZMs
PZMs
OFMMs
RHFMs
CHFMs
10
20
30
40
50
60
70
80
90
100
10 20 30 40 50 60 70 80 90 100
P
r
e
c
i

s
i
o
n

(
%
)
Recall (%)
COIL-100
RMs
CMs
ZMs
PZMs
OFMMs
RHFMs
CHFMs
10
20
30
40
50
60
70
80
90
100
10 20 30 40 50 60 70 80 90 100
P
r
e
c
i

s
i
o
n

(
%
)
Recall (%)
Noise
RMs
CMs
ZMs
PZMs
OFMMs
RHFMs
CHFMs
10
20
30
40
50
60
70
80
90
100
110
10 20 30 40 50 60 70 80 90 100
P
r
e
c
i

s
i
o
n

(
%
)
Recall (%)
Rotation
RMs
CMs
ZMs
PZMs
OFMMs
RHFMs
CHFMs
63

2.8.1. Discussion

In the above section, various radial moments are described along with their
properties. In our experiments, the performance of radial moments RMs, CMs, ZMs,
PZMs, OFMMs, RHFMs, and CHFMs is analyzed for retrieving images for several kinds
of databases. These databases consist of images representing numerous aspects of images,
such as distortion, partial occlusion, pose variation, rotation, noise effect, 3D objects, gray
scale, and binary images, etc. By evaluating the above R P

curves for all databases, we
observe that radial moments ZMs, PZMs, OFMMs, RHFMs, and CHFMs, are capable
enough to meet six principles set by MPEG-7. Among them, ZMs supersede rest of the
moments for image retrieval effectiveness. It is worth mentioning here that all the above
mentioned radial moments are highly invariant to rotation and provides 100% retrieval
rate for rotated images. Although, all moments represent good robustness to noise, PZMs,
RHFMs, and CHFMs can be recommended for best performance for noisy images. While
considering the computation perspective of radial moments, we observe that ZMs takes
least amount of CPU time for the same order of moments as compared to other moments.
OFMMs, RHFMs, and CHFMs take the highest and similar amount CPU time. Thus, the
above mentioned observations lead to the recommendation of ZMs for effective and
efficient image retrieval.
2.9. Performance evaluation of similarity measures
For evaluating the performance of different similarity measures a detailed
experimental analysis is performed. The objective of our experiments is to obtain evidence
that which similarity measure attains high performance and in what sort of circumstances,
it can be applied on. The performance of various similarity measures viz. Euclidean
distance, Manhattan distance, Bray-Curtis distance, Hellinger distance, Chebyshev
distance, Chi-square statistics and Cosine distance is analyzed on the image retrieval
system. Besides, a novel similarity measure referred to as Min-Max is proposed and
evaluated against other similarity measures. The region based descriptor angular radial
transform (ART) and contour based Fourier descriptors (FD) are used for extracting
features, which are described in the next chapter. The experimental study for assessing the
performance of similarity measures is given in Appendix B. After experimenting on
standard image databases, we conclude that
64

- The performance of similarity measures is almost similar for both descriptors ART
and FD. For MPEG-7 database, the performance of Euclidean, Manhattan, Bray-
Curtis, and Hellinger distances is similar and superior to that of Chebyshev,
Canberra, and Chi-square distances. However, the proposed Min-Max outperforms
all the similarity measures. For Kimia-99 database, the highest retrieval rate is
given by Min-Max and the lowest by Canberra distance. While considering BEP
of the similarity measures, it is apparent that Min-Max similarity measure provides
the highest retrieval rate. The top retrieval results reveal the superiority of both
Min-Max and Bray-Curtis similarity measures.
- The performance of Euclidean distance is slightly lower than Bray-Curtis. This is
due to the fact that in Euclidean distance, the values in each dimension are squared
before summations, which formulate more emphasis even on dissimilar features.
However, in Bray-Curtis similarity measure the distance value will never exceed
one because, in its equation, the numerator signifies the difference and
denominator normalizes that difference. Therefore, it could be used to avoid
scaling effect. The similar concept is adopted in Min-Max similarity measure. The
worst retrieval performance is given by the cosine similarity measure.
- While considering the computation complexity, Manhattan similarity measure is
the simplest to compute among others, whereas the cosine similarity measure has
the highest computation complexity.
- By considering the experimental results, we conclude that Min-Max, Bray-Curtis,
Euclidean, Manhattan, and Hellinger similarity measures can be used for effective
classification and in obtaining the improved image retrieval rate.

2.10. Conclusion

In this chapter, some fundamental tools for image retrieval, which are based on
shape have been elaborated. The hypothesis and properties of various radial moments are
described and discussed. It has been shown that the spectral features extracted by the
above mentioned moments are proficient to represent and describe shapes effectively. The
performance of all the radial moments is also analyzed over several databases, which
contain various sorts of images. It has been observed that all the radial moments provide
good retrieval accuracy. However, ZMs are observed to perform well in terms of accuracy
and efficiency.
65

Apart from that, various similarity measures are also described and discussed in
this chapter. The performance of similarity measures is evaluated and it has been observed
that Min-Max and Bray-Curtis similarity measures can be suggested for effective image
retrieval. Both Min-Max and Bray-Curtis provide normalized distances of the spectral
features of query and database images. Consequently, these measures eliminate the scaling
effect of large and small magnitude features and provide good retrieval results. Since the
databases used for image retrieval tests contain images of relevant and irrelevant classes,
precision-recall and Bulls eye performance, performance measures are used to evaluate
the image retrieval performance throughout the research work.

You might also like