Content Based Image Retrieval Using SVM for Relevance Feedback

jamaicacooperativeΤεχνίτη Νοημοσύνη και Ρομποτική

17 Οκτ 2013 (πριν από 3 χρόνια και 9 μήνες)

86 εμφανίσεις

INSTITUTE OF TECHNOLOGY, NIRMA UNIVERSITY, AHMEDABAD
-
382 481.25
-
27 NOVEMBER
, 2009

1



Abstract



This research
shows

an advantage of
interfacing Support Vector Machine (SVM) in
Relevance Feedback

(RF)
.

SVM
uses user’s
feedback to provide
interface between

q
uer
y

by
example

image
during
content based image
retrieval

(CBIR)
.
In CBIR,
images
are retrieved
from image database using color feature

but
semantic gap is generated betw
een color feature
and human

perception.
As
an
example, different
users have

different relevant images

for same
example image
, also same user has differ
en
t
relevant imag
es
at different time.
Th
e
s
e all

problem
s

are

solved by

Relevance Feedback.
During
retrie
val

process,
user s
elects
most relevant
images

and

SVM learnin
g results are used to
update
weights of preference
s

for relevant images.
Priorities are given to the posi
tive feedbacks that
have larger distances to hyper

plane determined by

support vectors.
Based on this SVM, new retrieval
images are displayed to user.
Experiment results
show

that
proposed (
RF using SVM)

approach
improvement
results

over CBIR

results
.



In
dex Terms

-

CBIR, Relevance Feedback, RF,
S
VM
.


I.

I
NTRODUCTION


Interest in potential of digital images
has increased enormously over the last

few years.
In today’s world digital images and videos have
become more effective ways of expression /
comm
unication and information capturing. They
are playing increasingly important role every
days life in modern information society as
architecture and engineering design, fashion
&
interior design, journalism and advertising,
medical diagnosis, geometric info
rmation system
and remote sensing, culture heritage, education
and trainin
g, entertainment, web searching,
military
etc.

In typical
,

content
-
based image retrieval
sys
tems
[1]

(Fig.

1
), visual contents of the images
in
d
atabase are extracted and described by

multi
-
dimensional feature
(color, texture, shape
features)

vectors.
F
eature vectors of images in
database form a feature database. To
retrieve
images, users provide
retrieval system with
example images.
S
ystem then changes these
examples into its internal
repres
entation of
feature vectors. S
imilarities distances between
feature vectors of query example and those of

images in
database are then calculated and
retrieval is performed with aid of an indexing
scheme.
I
ndexing scheme provides a
n efficient
way to
search for
image database. Recent
retrieval systems have incorporated users’
relevance feedback to
modify
retrieval

process in
order to generate perceptually and semantically
more meaningful retrieval results.


Bayesian network method,
nearest

neighbor
sea
rch method, Novel log
based Relevance
Feedback method and

Support Vector Machine
method are used for relevance feedback.
Support
Vector Machine method is best because of its
fast accessible and also high dimensional data,
which are not compressed.



Fig.
1. Diagram for Content Based Image Retrieval using
Relevance Feedback


Relevance Feedback learns
associations
between human perception and color feature
. In a

Content Based Image Retrieval Using
SVM for
Relevance Feedback

1
Darshana Mistry
,

2
Swati Jain

1
G
andhinag
a
r Institute of Technology,
Gandhinagar
,
2
Nirma Institute Of Technology
,
Ahmedabad,







2






NATIONAL CONFERENCE ON CURRENT TRENDS IN TECHNOLOGY,’NUCONE
-
2009


typical, relevance feedback system, given a set of
retrievals for an image query, user identifies

relevant and non
-
relevant examples. Based on
these examples, the similarity metric is modified
to re
-
compute

for next set of

retrievals and
displayed to

user.

II.

C
OLOR
F
EATURE

Color is first and most straightforward visual
feature for indexing and retrieval

of images[
5
].
They are usually robust in noise, resolution,
orientation and resizing.
Because of
their little
semantic meaning and its compac
t
representation, color feature

tend to be more
domain independent compared to other features.

The image retrieval
process can be divided in
two steps[
3
]:


Indexing for each image in

database a set or
a vector of features summarizing
. I
ts content
propert
ies is computed and stored in

metadata
database.


Retrieval given a query image
,

its features
a
re extracted and com
pared to
other
s in
database.
Database images are then ordered following a
similarity

criterion.

In color indexing, given a query image, goal is
to retrieve all images

whose col
or compositions
are similar to
color composition of query
image[
4
].Typically, co
lor

content is
characterized by color histograms, which are
compared using histogram distance measure. A
color histogram can be constructed in 3D color
space. This propose
d

method for extending use
of image histograms to characterize local and
global color

properties of images.

This method is based on two types of color
representations:


Color Descriptor to represent global color
features of images.


Color Descriptor Matrix to repre
sent spatial
color features of images
.

The color descriptor is defined to

be:

Color Descriptor

={

{
c
i,


p
i
}
, i
=1

M}
…….
(1)


Where M is total number of color clusters in
image, c
i

is an Arabic
number

corresponding

to
color,
p
i

is its percentage, and

p
i
=1.

Note that
M can vary

from image to image. First, most
dominant color in image is added and then

less
dominant color follows
.
As each pair {c
i
, p
i
} is
added,
color descriptor becomes more expressive
representation of color distribution of an image.
C
urrently, there are t
en colors in the definition of
color code book,

color code is changed as image.

Try to combine image quantization and
perceptual color model to represent spatial color
information in images and define 2
-
D vector
Color Descriptor Matri
x.

Color

Descriptor

Matrix


=
{Ci,j
,i=0...N,j
=0...N}
……...
(2)


In
order to create this structure
whole image is
divided into NxN (where N=4,

8, 16, 32) equal
parts. This matrix stores
dominant colors for
image blocks.
O
riginal images were NxN
quantized and were represented as NxN blocks
(or sub images).

Original

images

are 16X16
quantized and represented as 16X16 color blocks
(as

Fig.
2).



Fig.
2
: The

process of color descriptor matrix


Q
uantized image
[2]

is divided i
nto 5
x
5 matrix
and 16 bin histogram is calculated for each sub
image and most dominant color is new
replacement of small block and hence 125
x
83
images are reduced to 25
x
16.

III.

S
UPPORT
V
ECTOR
M
ACHINE

Support Vector Machines (SVM) is an
approximate implementati
on of structural risk
minimization (SRM) principle
[7]
. It creates a
classifier with minimized

Vapnik
-
Chervonenkis
(VC
)

dimension
. SVM minimizes an upper
bound on generalization error rate.
E
rror rate is
INSTITUTE OF TECHNOLOGY, NIRMA UNIVERSITY, AHMEDABAD
-
382 481.25
-
27 NOVEMBER
, 2009

3

bou
nded by sum of the training

vectors
.
Consider
prob
lem of separating set of training vectors
belonging to two

classes, e.g., image retrieval
problem, +1 denotes positive example,
-
1
denotes negative example.


{x
i
, y
i

}
N
i=1
, y
i

=+ 1/
-
1……………………..(3)

Where

is an input pattern, e.g
., feature
vector in image retrieval, for i
th

example and
y
i

is
label. If two classes are linearly separable, hyper
plane that does separation is:

w
T


+ b = 0
………………………………
(4)

Where
an input vector, w is

a weight vector,
and b is a bias. The goal of a

sup
port vector
machine is to find
parameter w
o

and b
o

for a
optimal hyper plane to maximize the distance
between the hyper plane and the closest data
point:

w
o
T


+ b
o

≥ 0 for


y
i

= +1
……………….
(5)

w
o
T


+ b
o

≤ 0 for

y
i

=
-
1
………………..
(6)


For a given w
o
, and b
o
, distance of a point

from optimal hyper plane

defined in equation (5)
and (6) is


d
(w
o
, b
o
,
) =
|
w
o
T


+ b
o
| / ||
w
o
||
………
(7)


Fig.

3:

Illustration of the idea of an
optimal hyper

plane for
line
arly separable patterns and definition of

distance



A linear separable example in 2D is illustrated
in Figure 3. If two
classes

are non
-
linear
ly
separable, input vectors
should be nonlinearly
mapped to a

high

dimensional feature space by
an in
ner
-
product kernel function
.

This hyper

plane is optimal in sense of being a
maximal margin classifier

with respect to

training
data.

D
istance from hyper

plane determined

by
support vectors can be used to measure how
much an example belonging to one class is
different from other class. This motivates us to
use SVM for automatically generating preference
weights for relevant images.

IV.

EXPE
RIMENTAL
RESULTS


Using
of MATLAB

2007b and Post
gre
SQL
8.3
, experimental

results are as shown

below:


First Query is given by example image which is
select
ed

from l
ist box or allocated path, i
.e. user
select cloudy image as Fig. 4. Based on this
exampl
e image, total 36 image
s

using
color
feature are retrieved.
Here,

user
has given 10
positive
feedba
ck
s
for relevant image
s
.

Based on
these relevance feedback images (
SVM

is used)
,
total relevant images

are 21

as user perception
from 31 images.
T
otal
cloudy

images
are 30
in a

100 images

database.
C
olor based image
retrieval’s relevant image
s

efficiency is (10/30)
33.33%

(see Fig. 5(a)(b)(c))

and relevant image
s

based on SVM is (21/30) 70%

(see Fig
6(a)(b)(c))
.





Fig. 4. Query by example


Different users

select same 107.jpg example
image and results are as
TABLE I and Fig. 7.


Efficiency of relevant images using SVM is also
better than content based relevant images in 1000
images database (119.jpg, 441.jpg, 464.jpg,
762.jpg examples of TABLE II). Survey o
f 25
users, average relevant images efficiency using
SVM is 78.6% which is better than content based
relevant image using color feature (51.13%) as
TABLE II.


V.

C
ONCLUSION

The experiments were conducted with color
database images. Images retrieved based on c
olor
4






NATIONAL CONFERENCE ON CURRENT TRENDS IN TECHNOLOGY,’NUCONE
-
2009


feature. But the semantic gap in user perception
and obtained output is quite usable.


(a)
(b)


( c )


Fig. 5
(a), (b) (c)
Images are retrieve
d

based on color feature


(a)


(b)


( c )

INSTITUTE OF TECHNOLOGY, NIRMA UNIVERSITY, AHMEDABAD
-
382 481.25
-
27 NOVEMBER
, 2009

5

Fig. 6.
(a),(b) and (c)

Images are retrieve based on relevance

feedback using SVM

TABLE

I






D
IFFERENT USER SELECT

SAME
107.
JPG EXAMPLE IMAGE




Fi
g. 7. Graph for different user


same example image


Using of relevance feedback with S
VM, results
are more efficient as user perception. SVM
classification can be even better if the feature
vector used in more relevant to images.

VI.

R
EFERENCES

[1]

D. F. Long, D. H. Zhang, and P. D. D. Feng,

Fundametal of content based image retrieval,” reserch
mi
crosoft, 2003.

[2]

S. Jain and S.N.Pradhan, “Enhancement of color imag
e
retrieval capabilitites:
function of color with
texture(optimized),” NUCONE 2007,Nirma University,
December 2007.

[3]

I. Valova and B. Rachev, “Image databases an
approach for image segme
ntation

and color reduction
analysis and synthesis,”
International Conference on

Computer Systems and Technologies
-

CompSysTech2003
,
.

[4]

I. Valova and B. Rache, “Retrieval by color features in
image databases,” International Conference on
Computer Systems

and Techn
ologies
-

CompSysTech2002.




TABLE
II

AVERAGE EFFIECNCEY O
F RELEVANT IMAGES

(%)

BASED ON
COLOR FEATURE AND
RF(SVM)



[5]

I. Valova, B. Rachev, and M. Vassilakopoulos,
“Optimization of the algorithm for image retrieval by
color features,” Intern
ational Conference on Computer
Systems and Technologies
-

CompSysTech, 2006.

[6]

I. Valova and B. Rachev, “Image databases
-

an
approach for image segmentation and color reduction
analysis and synthesis,” International Conference on
Computer Systems and Techn
ologies
-

CompSysTech2003.

[7]

Q. Tian, H. Pengyu
, and H. T
homas
, “Update relevant
image weights for content
-
based image retrieval using
support vector machines,” 2000 IEEE International
Conference on volume 2, 2002
.

[8]

O. T
akashi
. and M. H
iroshi

and Y. S
eiji
,

“Relevance
feedback document retrieval using support vector
machines,” 2004 IEEE International Joint Conference
on Volume 4, pp. 1359


1364, July 2004.

[9]

C.
H.

Hoi and
M.

Lyu,

“Group
-
based relevance
feedback with support vector machine ensembles,”
ICPR 20
04, 17th International Conference on Volume
3, page no. 874


877, August 2004.


[10]

C. H. Hoi, C.H. Chan,
K. Huang,

M. Lyu, and I.King,

“Biased support vector machine for relevance
feedback in image retrieval,” 2004 IEEE International
Joint
Conference

on

Volume 4, pp. 3189


3194, July
2004.