Computer Science Section

41


Relevance Feedback in Content Based Image Retrieval: A Review

1
Pushpa B. PATIL,
2
Manesh B. KOKARE
1
B.L.D.E.As, College of Engineering and Technology, Bijapur-586103, India
2
S.G.G.S. Institute of Engineering and Technology, Nanded-431606, India
1
pushpa_ms@rediffmail.com,
2
mbkokare@sggs.ac.in

AbstractThis paper provides an overview of the technical
achievements in the research area of relevance feedback (RF) in
content-based image retrieval (CBIR). Relevance feedback is a
powerful technique in CBIR systems, in order to improve the
performance of CBIR effectively. It is an open research area to
the researcher to reduce the semantic gap between low-level
features and high level concepts. The paper covers the current
state of art of the research in relevance feedback in CBIR,
various relevance feedback techniques and issues in relevance
feedback are discussed in detail.

Keywords: Relevance feedback, long-term learning, short-term
learning, image retrieval, content-based image retrieval, and
semantics

I. INTRODUCTION

Recently, there is a rapid growth of digital image data on
the Internet and in digital libraries. The advent of Internet has
made information sharing and access easier. Internet users are
indulging into information exchange. Retrieving information
from the World Wide Web has become a common practice.
However with the day by day increase in size of the web,
abundant information introduced heterogeneity of this
information makes classical information retrieval techniques
ineffective. Searching and retrieving information as desired
has become a serious challenge.
Information sharing has increasingly become a common
phenomenon among the users of todays high speed networks.
Advancements in technology, enables a variety of different
types of information available. However this heterogeneity
surely challenges technology to provide efficient ways for
accessing, sharing, storage of such heterogeneous information
over the networks and databases. Due to advancements in the
digital photography technology, large storage capacity and
high speed networks, storing large amounts of high quality
images has become possible. Digital images find a wide range
of applications in field of medicine, science (medical and
scientific images), at virtual museums and galleries, military
and security purposes, and personal photo albums etc. while
dealing with this sort of information like organizing and
searching large volumes of images in databases, users can
have difficulties, as the current commercial database systems
are designed for textual data, it is not well suited and
compatible for digital images. Therefore there is a need for an
efficient way for image retrieval. In order to cater to this
need, researchers have tried extending the current information
retrieval (IR) techniques that are used in text retrieval to the
area of the image retrieval.
There are different ways to retrieve the images in CBIR.
Ritendra Datta et al [1], Rui and Huang [2], Smeulders et al
[3] and Kokare et al [4] had presented comprehensive and
recent extensive literature survey on content based image
retrieval. The oldest method is text annotation to images in
the database. Image annotation is tedious task. Because, it is
practically impossible to annotate all the images in the
databases. Second it is also very difficult to label the same
annotations to the same image by different users. To address
such significant limitations, researchers have turned their
attention to content-based image retrieval. In CBIR systems,
low level image features are extracted based on visual content
such as color, shape and texture. Which are represented by
feature vectors instead of a set of keywords. However, big
challenge in CBIR is the semantic gap between the low level
features and high level concepts. In order to reduce the gap
between the low level features and high level concepts,
relevance feedback was introduced into CBIR[5],[6].
Recently, many researchers began to consider the RF as a
classification or learning problem. That is a user provides
positive and/or negative examples, and the systems learn from
such examples to separate all data into relevant and irrelevant
groups. Hence many classical machine learning schemes may
be applied to the RF such as, decision tree learning [7],
Bayesian learning [8], [9], support vector machines [10],
boosting [11] and so on. Another challenge in CBIR systems
is multidimensional indexing. In CBIR systems, the visual
content of image features is high-dimensional numerical data.
So it is difficult to manage these data with traditional database
systems, because these systems are designed for text data and
low-dimensional numerical data. Therefore many researchers
have proposed systems for indexing high-dimensional data in
CBIR systems.
The main contributions of this paper are summarized as
follows. Firstly, in this paper, we have provided brief survey
of CBIR. Secondly, we have discussed how relevance
feedback in content-based image retrieval is used and its
current state of the art. Finally, future directions in relevance
feedback in CBIR are also suggested.
The paper is organized as follows. In section II, we discuss
the system architecture. In section III, we review the current
state of art of research on relevance feedback in CBIR, and

Journal of Applied Computer Science & Mathematics, no. 10 (5) /2011, Suceava

42
different RF techniques are discussed. In section IV,
challenges in RF are discussed. Finally, conclusion is
presented in section V.

II. SYSTEM ARCHITECTURE

Fig. 1 shows the general scheme of image retrieval from a
database using relevance feedback. The basic idea of
relevance feedback is to shift the burden of finding the right
query formulation from the user to the system. In order to
make this true, the user has to provide system with some
information, so that system can perform well in answering the
original query. To retrieve the image from the database, we
first extract feature vectors from images (the features can be
shape, color, texture etc), then store feature vectors into
another database for future use. When given query image, we
similarly extract its feature vectors, and match those features
with database image features. If the distance between two
images feature vectors is small enough; we consider the
corresponding image in the database similar to the query.
The search is usually based on similarity rather than on
exact match, and the retrieval results are given to the user.
Then user gives the feedback in the form of relevance
judgments expressed over the retrieval results. The relevance
judgments evaluate the results based on a three value
assessment. These three values are relevant, non-relevant and
dont care. Relevant means the image relevant to the user,
non-relevant means the image is definitely not relevant, and
dont cares mean the user does not say anything about the
image. If the user feedback is relevant, then feedback loop
stops otherwise it continues until user get satisfied with
results
In Fig. 1, the block diagram consists of following main
blocks like image database, feature extraction, similarity
measure, user feedback, and feedback algorithm. The function
of each block is discussed below. Where U/R is the user
satisfaction or result remains same.

A. Feature Extraction
Feature extraction involves extracting the meaningful
information from the images. So that it reduces the storage
required and hence the system becomes faster and effective in
CBIR. Once the features are extracted, they are stored in the
database for future use. The degree to which a computer can
extract meaningful information from the image is the most
powerful key to the advancement of intelligent image
interpreting systems. One of the biggest advantages of feature
extraction is that, it significantly reduces the information
(compared to the original image) to represent an image for
understanding the content of that image. There has been
tremendous work on different approaches to the detection of
various kinds of features in images.


Fig. 1. System Architecture

These features can be classified as global features and local
features. The most commonly used features are color, texture,
and shape. They are all application independent.
1)Global features: Global features should be calculated
over the entire image. For example, average gray level, shape
of intensity histogram etc. The advantage of global extraction
is its high speed for both extracting features and computing
similarity. Specifically, they can be oversensitive to location
and hence fail to identify important visual characteristics. To
increase the robustness to spatial transformation, we can go
for local feature extraction
2) Local features: In global features, the features are
computed from the entire image. However, these global
features cannot handle all parts of the image having different
characteristics. Therefore, local features of the image are
necessary. These features can be calculated over the results of
image segmentation and edge detection algorithms, that is,
they are all based on the part of the image with some special
properties.
3) Salient points: In local feature computation, the feature
extraction of the image is limited to a subset of the image
pixels; the interest points, the set of interest points are called
salient points. The salient points are points of high variability
in the features of the local pixel neighborhood. Many CBIR
systems extract salient points[12][13]. In [14], they defined
localized content-based image retrieval as a CBIR task, where
the user is only interested in a portion of the image, and the
rest is irrelevant. For example, we can refer to some local
features as image primitives; circles, lines, texels (elements
composing a textured region) other local features; shape of
contours etc.

B. Similarity Measure
In similarity measure, the query image feature vector and
database image feature vector are compared using the
distance metric. The images are ranked based on the distance
value. It is proposed in [15], the detailed comparison of nine
different metrics such as Manhattan, weighted mean-variance,

Computer Science Section

43
Euclidean, Chebychev, Mahanobis etc distance for texture
image retrieval with empirical evaluation. They found that
Canberra and Bray-Curtis distance metrics performed
exceptionally well than all other distance metrics.

C. User Feedback
After obtaining the retrieval results, user provide the
feedback as to whether the results are relevant or non-
relevant. If the results are non-relevant the feedback loop is
repeated many times until the user is satisfied. The typical
scenario for relevance feedback in CBIR is as follows
Algorithm: Typical scenario for relevance feedback in
CBIR
Begin
Obtain the initial retrieval results of CBIR
Repeat until user satisfaction or result remains same
From user interaction, obtain the feedback from
the users on prior results. Feedback is in the
form of relevant or irrelevant to request.
If results found to be not satisfied
Learn the system through a feedback algorithm
and hence results are refined
End repeat
End

III. RELATED WORK IN RELEVANCE FEEDBACK

The concept of relevance feedback was introduced into
CBIR from the concept of text-based information retrieval in
the 1998s [6] and then has become a popular technique in
CBIR. In 1998, Chang, et al.[16], proposed a framework,
which allows for interactive construction of a set queries
which detect visual concepts such as Sunsets. In 2001,
Scloroff, et al. [17], describes the first WWW image search
engine, which focused on relevance feedback based
improvement of the results. In their initial system, a relevance
feedback was used to guide the feature selection process; it
was found that the positive examples were more important
towards maximizing accuracy than the negative examples.
In 2001, Rui and Huang [18], compared heuristic with
optimization based parameter updating and found that the
optimization based method achieves higher accuracy. In
2001, Chen, et al.[19], described a one class SVM method for
updating the feedback space which shows substantially
improved results over previous work. In 2001, Guo, et al.[20],
performed a comparison between AdaBoost and SVM and
found that SVM gives superior retrieval results. In 2002, He,
et al.[21], used both short term and long term perspectives to
infer a semantic space from users relevance feedback for
image retrieval. The long term perspective was found by
updating the semantic space from the results of the short term
perspective. In 2003, Dy, et al.[22], proposed a two level
approach via customized queries and introduced a new
unsupervised learning method called feature subset selection
using expectation-maximization clustering. The proposed
method doubled the accuracy for the case of a set of lung
images. In 2004, Tieu and viola [23], proposed a method for
applying the AdaBoost learning algorithm and noted that it is
quite suitable for relevance feedback due to the fact that
AdaBoost works well with small training sets. In 2005, Yin,
et al.[24], found that combining multiple relevance feedback
strategies gives superior results as opposed to any single
strategy.
In [25], a unified log-based relevance feedback network for
integrating log data of user feedback with regular relevance
feedback for image retrieval was proposed. Here the
framework first compute the relevance functions on the log
data of user feedback and then combines the relevance
information with regular relevance feedback for the retrieval
task. In order to address the noisy log data problem in real
world applications, a novel learning algorithm to solve the
log-based relevance feedback problem was proposed. The
proposed algorithm, named Soft Label support vector
machines, is based on the solid regularization theory.
In [26], a relevance feedback image retrieval method based
on PCA image classification was proposed. From the
experiment it could be seen that more accurate classification
result can be obtained by using PCA clustering algorithm.
The proposed method can not only improve retrieval
precision greatly but also reduces retrieval time and
complexity. In [27], proposed the SSAIRA method was
proposed, which addresses three special issues of relevance
feedback i.e small sample size, asymmetrical training sample,
and real time requirement. It was shown that the proposed
SSAIRA method is superior to some existing methods. This
method employs semi-supervised learning and active learning
simultaneously, which is beneficial to the improvement of the
retrieval performance.
A. Relevance Feedback Approaches
In relevance feedback-based approaches, a CBIR system
learns from feedback provided by the user. Learning in CBIR
systems is categorized into short-term learning, and long-
term learning [21]. Query refining using relevance feedback
has gained much attention in the research and development of
content-based image retrieval systems. Most of the researches
have focused on query tuning in a single retrieval session.
This is commonly known as intra-query learning or short-
term learning. In contrast, inter-query, also known as long-
term learning is strategy that attempts to analyze the
relationship between the current and past retrieval sessions.
1) Short-Term Learning Approach: In short-term learning,
only the feedbacks for the current search session are used in
the learning algorithm, and image features are the primary
source of data. The main challenge in this approach is to find
the best combination of image features that presents the users
query. Such optimum set of feature can include features that
capture similarities between positive images, or features that
discriminate positive examples from negative ones. Hence
many classical machine learning algorithms may be widely
used in short-term learning, which include support vector
machines [10], [29]-[30], Bayesian learning [8]-[9], boosting

Journal of Applied Computer Science & Mathematics, no. 10 (5) /2011, Suceava

44
[11] and feature weighting, discriminant analysis so on.
However, the process of learning in short-term approach is
very difficult task, due to; firstly, the size of the training data
is much smaller than dimension of feature space. Secondly,
there is too much imbalance between the users feedback.
Lastly, the process of learning is online, which takes more
real time.
2) Long-Term Learning Approach: Long-term learning
methods can achieve better retrieval precision compared to
traditional RF techniques. We can use long-term learning to
overcome the difficulties like incapability of capturing
semantics scarcity and imbalance of feedback examples, and
lack of memory mechanism etc [28]. Actually, the concept of
long-term learning in CBIR is adopted from the work of
collaborative filtering. Long-term learning approaches utilize
the feedbacks collected during prior retrieval iteration. It is a
cumulative process for collecting fast feedbacks and which
are stored in the form of matrix. A matrix stores the labels
provided by user for each image in every iteration. As usually
the size of search history matrix is large, statistical models
and approaches such as principal component analysis and
latent semantic analysis are popular in long-term learning
approaches. However, there are problems in long-term
learning methods.
Limitations:
1) Firstly, this is approach was shown to be unsuitable for
applications where images were frequently added or removed.
A better approach is to use the feature vector model to
analyze the inter-query relationship. In [31], one class
Support Vector Machine (1-SVM) was used to calculate the
query point for the search session. The system only stores the
positive centroid of each query session. The inter-query
similarity measure is performed by calculating the distances
between the positive centroid points of the present and past
query sessions.
2) Secondly, the sparsity of memorized feedback
information. The quality of long-term learning depends
strongly on the amount of user log that the system has stored
so for. Because of the lack of interactions and large database,
it is not easy to collect sufficient log information. In [28],
introduced a feedback knowledge memory model to
accumulate the previous users preferences. They also
presented a learning strategy to predict hidden semantics
using the memorized information, which is able to reduce the
limitation of user log sparsity to a certain extent. Combining
these two is called memory learning.
3) Lastly, the other problem is that most long-term learning
approaches only recommend the memorized semantic
knowledge to users but lack a learning ability to predict
hidden semantics in terms of acquired semantics. In [28], to
address this issue, they proposed an annotation propagation
scheme on a semantic level by the memory learning.

B. Query Refining Methods in Relevance Feedback
Recently, there are several relevance feedback methods
proposed. In [32], the query point movement and re-
weighting method are used to capture the geometrical
information of relevant images in a feature space. In [8],
probabilistic models are estimated to predict relevant
images. In [19], machine learning methods are used to solve
the classification problem for two classes, so that we can
classify the relevant and irrelevant images. In [10], the query
concept is learned by means of active learning algorithms,
such as SVM active and MEGA. The early relevant feedback
schemes for CBIR have been mainly adopted from document
retrieval researches and can be classified into two approaches.
Query modification and re-weighting. Both have been built
based upon vector model in information retrieval theory [33].
It allows users to refine the query representation. A user may
start from a query object that approximately captures his
information need. In each feedback iteration, the system
modifies the representation of the query to a more suitable
representation. Query modification can be achieved using
either of two approaches; query point movement and updating
weight vector.
1) Query Point Movement: It is essentially tries to
improve the estimate of the ideal query point by moving it
towards good example points and away from bad example
points. Query is represented by a single point in a feature
space and refinement process attempts to move that point
toward the direction where relevant points were located. The
frequently used technique to iteratively improve this
estimation is Rocchios formula given bellow [33]. This is
technique implemented in the MARS system [5].


( ) ( )

=

=
+
+ =
+
J
rel non
j
Y
j
n N
J
rel
j
X
j
n N
q
n
q
n
_
1 1
1

(1)

Where
q
n
is the query point for n
th
round of the search
cycle; parameters , and suitable constants, also
commonly known as the weight parameters ;
J rel
is the
number of relevant images in X
j
and J
l
e r non_

is the total
number of non-relevant images in Y j . Depending on the
nature of the data samples, the parameters and can be
adjusted to be more biased towards one sample group. It
should be noted that the negative sample may totally ignored
if variable is set to zero, and the history of the query point
can also be disregarded by setting variable to zero.
2) Updating Weight Vector: Query weighting changes
the relative weights of different features in the query
representation. The updating weight vector mechanism allows
the system to learn the users interpretation of similarity
/distance function. The central idea behind the re-weighting
method is very simple and intuitive. Since each image is
represented by an N dimensional feature vector, it can be
viewed as a point in an N dimensional space then the basic
idea is to enhance the importance of those dimensions of a

Computer Science Section

45
feature that help in retrieving the relevant images and reduce
the importance of those dimensions that hinder this process.
The weighting method [34] associates larger weights with
more important dimensions and smaller weights with less
important ones. An ideal query vector for each feature i is
described by the weighted sum of all positive feedback
images as


1
T
i
n
i
j
j
Q
y

=
=

(2)

where
y
i
is the
k
n
i
X
( k
i
is the length of feature i) training
example matrix for feature i obtained by stacking the n
positive feedback training vectors
x
i
+
into a matrix . The n
element vector
]
...
1
[ n
=
represents the degree of
relevance (to the query) of each of the n positive feedback
images, which can be determined by the user at each feedback
iteration. The system then uses
q
i

as the optimal query to
evaluate the relevance of the images in database. This strategy
is widely used by many other image retrieval and relevance
feedback systems[ 25]-[26].
This approach is to cover the target images by increasing the
value, i.e. the weight, of the important features, while
reducing the value of non-relevant feature. If a vector is used
to represent the feature space, then the re-weighting scheme
merely becomes parameter adjustment along the line of
independent axis in the feature space. As an example, the
general form of the Euclidean distance measure between a
feature vector extracted from an image and the positive
centroid of an image group becomes

( ) ( )

=
=
K
k i
c i k c i
k f k f f f d W ) (
,
(3)

Where } ,... { 3 , 2 , 1 K k = is the length of the vector. Vectors
f
i
and f
c
are the new feature vector and the positive
centroid of an image group respectively. Determining the
weights of the feature vector has been a focus of many
research reports.
3) Probabilistic Approach: The probabilistic approach
has been one of the most commonly used models in text
document retrieval systems. In information retrieval systems,
this probabilistic model may be thought of as a multistage
random experiment, where each image in the database is
associated with an estimated probability of certain condition.
During each session of retrieval, the probabilities associated
with each database image change according to the results
obtained from previous sessions. It can be described in the
form of conditional probabilistic model.
In CBIR, PicHunter by Cox et al. [35] was the first to use a
probabilistic model in relevance feedback to extract the
information from the retrieval pattern. The approach used in
the PicHunter system was based on modeling of the history of
the system image display patterns,
D
t
, and the users
feedback,
A
t
, to each display. Using the Bayes theorem and
the sequences of the retrieval sessions, Cox et al. expressed
their probability formula as

=

=

=

=

=
= =
n
j
H
t
T
i
T P H
t
D
t
T
i
T A
t
P
H
t
T
i
T P H
t
D
t
T
i
T A
t
P
H
t
T
t
T P
1
)
1
| ( )
1
,
,
| (
)
1
| ( )
1
,
,
| (
) | (
(4)

Where
H
t
is the history of the previous retrieval
iterations, consisting of sequences of display patterns and
users feedbacks. The above expression may be viewed as an
incremental process and is designed to estimate the similarity
of an image to the users target image using his/her image
labeling patterns. This is called as the user model. In this
model, the estimation of the users judgment of the similarity
of images is based on the relative distance measure between
the images in the database and the user labeled images.
Bayesian estimation methods have been used in the
probabilistic approach for relevance feedback. Cox et al. [35]
used Bayesian learning to incorporate user feedbacks to
update the probability distribution of all the images

in the
database. They consider the feedback examples as a sequence
of independent queries and try to minimize the retrieval error
by Bayesian rules. That is, given a sequence of queries, they
try to minimize the probability of retrieval error as

)}
1
...
1
| ( ) | ( {
max arg
)
1
...
| (
max arg ) (
xt x
i y
P
i y xt P
xt xi
i y
P x g
i
i

= = =

=
=
(5)
Where } 1
...
{ xt xt is a sequence of queries (feedback
examples) and )
1
...
1
| ( xt x i y
P
= is a prior belief about
the ability of the i
th
image class to explain the queries.
Yin et al. [24] investigated that one of the drawbacks of the
Bayesian inference approach is that it requires more feedback
iterations or sample data to accurately approximate the
probability distribution of the samples, but this is not often
available in real time retrieval systems. Thus they suggested
that this technique should be integrated with other approaches
to gain more accurate results with less feedback iterations.
They proposed a framework for integrating the probability
approach together with point movement and re-weighting
approaches.

IV. CHALLENGES IN RELEVANCE FEEDBACK

In relevance feedback, the learner must use the training
data, i.e. the images marked by the user during subsequent
feedback rounds, and in order to estimate the target of the

Journal of Applied Computer Science & Mathematics, no. 10 (5) /2011, Suceava

46
user. The task of the learner is particularly difficult in the
context of relevance feedback for several reasons [27], [29],
[36], which are given below. The amount of training data is
very small, usually much lower than the number of
dimensions of the feature space. There are usually much
fewer positive examples than negative examples. The learner
must have a low sensitivity to this imbalance in the training
set or some corrective must be found. The target class may
have a rather complex shape or even several, rather
disconnected modes. Together with the fact that training data
is scarce, this can severely limit the generalization we can
expect. To preserve interactivity, both learning from the
training examples and the evaluation of the remaining
images according to the selection criterion must be very fast.
The computation cost can then be a very important criterion
in the choice of learning method. Challenges in RF CBIR are
as follows:
1) Unable to extract high level semantics: Most RF
techniques in CBIR, it is difficult to extract high-level
semantics of images when only low-level features are used in
RF. However, it works well for textual information retrieval.
Because retrieval is based on keywords but not on low-level
features.
2) Scarcity and Imbalance of feedback samples: Every
users are not wish to go through more number of feedback
iterations to get the best results. So, the number of feedback
samples labeled by the users during a RF session is far
smaller than the dimension of feature space. Due to such
small training data, most of the machine learning algorithms
cannot give exciting results. In addition, the number of
labeled negative samples is usually greater than the number of
labeled positive samples. As pointed out in [27], [36], the
imbalance of training data always makes classification
learning less reliable. Thus the small sample training data,
especially positive sample definitely limits the accuracy of RF
3) Real time processing: The process of learning in RF is
online and hence for every feedback iteration both training
and testing has to perform. Therefore it takes much real time.
A reasonable way to address this issue is to adopt efficient
image representation and storage structures such as the
hierarchical tree structure etc.

V. CONCLUSION

Relevance feedback is a powerful technique in CBIR
systems, in order to improve the performance of CBIR
effectively. It is an open research area to the researcher to
reduce the semantic gap between low-level features and high
level concepts. In this paper, detailed survey of relevance
feedback in CBIR is given. Various RF techniques and issues
are discussed in detail. For future research direction of RF in
CBIR are suggested.

REFERENCES

[1] Ritendra Datta, Dhiraj Joshi, Jia Li, and James Z. Wang., Image
Retrieval: Ideas, Influences, and Trends of the New
Age, ACMComputing Surveys, Vol. 40, No. 2 article 5,
pp.5:1-5:60, Apr.2008.
[2] Rui, Y., Huang, T. S., Chang, S. F., Image Retrieval: Current
Techniques, Promising Directions and Open Issues, J. Visual
Comm. and Image Representation, vol.10, pp. 39-62, Jan. 1999
[3] Smeulders, A. W. M., Worring, M., Santini, S., Gupta, A., Jain
R., Content based Image Retrieval at the End of the Early
Years, IEEE Trans. Pattern Anal. Machine Intell. 22(12),
pp.1349-1380, 2000
[4] Kokare, M., Chatterji, B. N., Biswas P. K., A Survey on Current
Content-based Image Retrieval Methods, IETE J. Res.
48(3&4), pp. 261-271. 2002
[5] Rui, Y., Huang, T.S., and Mehrotra,S. Content-based Image
Retrieval with Relevance Feedback in MARS, in Proc. IEEE
Int. Conf. on Image proc., 1997
[6] Rui, Y.; Huang, T.; Ortega, M.; Mehrotra,S. Relevance
Feedback : A Power Tool In Interactive Content-Based Image
Retrieval, IEEE Transactions on Circuits and Systems for
Video Technology , Vol. 8(5), pp. 644-655, 1998.
[7] S. D. MacArthur, C. E. Brodley, and C. R. Shyu, Relevance
Feedback Decision Trees in Content-based Image Retrieval
in Proc. IEEE Work-shop Content-based Access of Image and
Video Libraries, pp.68-72, Jun 2000.
[8] I. J. Cox, M.L. Miller, T.P. Minka, T.V.Papathomas, and
P.N.Yianilos, The Bayesian Image Retrieval System,
PicHunter: Theory, Implementation and Psychophysical
Experiments, IEEE Trans. on Image Processing, Vol.9, Issue
1, pp.20-37, Jan. 2000.
[9] Z. Su H. Zhang, S. Li, and S. Ma, Relevance Feedback in
Content-based Image Retrieval: Bayesian framework, Feature
Subspaces, and Progressive Learning, IEEE Trans. Image
Process, vol. 12, no. 8 , pp. 924-936, Aug 2003.
[10] Tong S. and Chang E., Support Vector Machines Active
Learning for Image Retrieval in Proc. ACM Multimedia, 2001.
[11] K. Tieu and P. Viola, Boosting image retrieval in Proc. IEEE
Conf. Computer Vision Pattern Recognition, pp. 228-235, Jun.
2000.
[12] Q. Tian , N. Sebe, M.S. Lew , E. Loupias, and T. S. Huang,
Image Retrieval Using Wavelet-Based Salient Points, J.
Electronic Imaging, special issue on Storage and Retrieval of
Digital Media, vol. 10, no.4 pp. 849-935, 2001.
[13] J. Wang, H. Zha, and R. Cipolla, Combining Interest Points
and Edges for Content-based Image Retrieval in Proc. IEEE
Intl Conf. Image Processing (ICIP 05), pp. 1256-1259, 2005.
[14] Rouhollah Rahmani, Sally A. Goldman, Hui Zhang, Sharath
R. Cholleti, and Jason E. Fritts, Localized Content-Based
Image Retrieval, IEEE Transactions on Pattern analysis and
Machine Intelligence, Vol. 30, No. 11, pp. 1902-1912, Nov.
2008.
[15] Manesh Kokare, B.N. Chatterji and P.K Biswas, Comparision
of Similarity Metrics for Texture Image Retrieval in IEEE
TENCON, pp. 571-575, 2003.
[16] Chang, S F, Chen. W., and H. Sundarm, Semantic visual
template: Linking Visual features to semantics in proceeding
of the IEEE International Conference on Image Processing, pp.
531-535, 1998.
[17] Sclaroff, S., La Casicia, M., Sethi, S., and Taycher, L., Mix
and Match Features in the ImageRecover Search Engine in
Principles of Visual Information Retrieval, M.S. Lew, Ed.
Springer-Verlag, London, pp. 259-277, 2001.

Computer Science Section

47
[18] Rui. Y and Huang, T. S, Relevance Feedback Techniques in
Image Retrieval in Principles of Visual Information Retrieval,
M. S. Lew, Ed. Springer-Verlag, London, pp. 219- 258, 2001.
[19] Chen, Y. ,Zhou , X.S., and Huang, T.S. ,One class SVM for
Learning in image Retrieval in proceeding of IEEE
International Conference on Image Processing, pp. 815-818,
Oct. 2001.
[20] Guo, G., Zhang, H.J., and Li, S. Z., Boosting for Content-
Based Audio Classification and Retrieval: An Evaluation, In
Proceedings of the IEEE Conference on Multimedia and Expo,
Tokyo, Japan, August 2001.
[21] He, X., Ma, W. Y., King , O. Li, M., and Zhang, H., Learning
and Inferring a Semantic Space from Users Relevance
Feedback for Image Retrieval In proceedings of the ACM
Multimedia , pp. 343-347 , 2002, New York.
[22] Dy, J.G., Brodley, C.E., Kak, A., Broderick, L.S., and Aisen,
A.M., Unsupervised Feature Selection Applied to Content-
Based Retrieval of Lung Images, IEEE Transactions on
pattern Aanalysis and Machine Intelligence, 25(3), pp. 373-
378, 2003
[23] Tieu, K.and Viola, P.,Boosting Image Retrieval, International
Journal of Computer vision, 56(1), pp. 17-36, 2004.
[24] Yin, P.Y., Bhanu, B., Chang, K.C., and Dong, A., Integrating
Relevance Feedback Techniques for Image Retrieval using
Reinforcement Learning, IEEE Transactions on Pattern
Analysis and Machine Iintelligence, 27(10). pp.1536-1551,
2005.
[25] Steven C.H., Hoi, Michael R.Lyu, Rong Jin, A Unified Log-
Based Relevance Feedback Scheme for Image Retrieval,
IEEE Transactions on Knowledge and Data Engineering, Vol.
18, No. 4, pp. 509-524, April 2006.
[26] Ming Li, Zhi-Yun Liu, Jain-kun Wang, An Adaptive
Relevance Feedback Image Retrieval Method with Based on
Passibilistic Clustering Algorithm in Proceedings of Sixth
International Conference Intelligent Systems Deign and
Applications, 2006.
[27] Zhi-Hua Zhou, Ke-Jia Chen, and Hong-Bin Dai, Enhanced
Relevance Feedback in Image Retrieval Using Unlabeled
Data, ACM Journal, pp. 1-25, 2006.
[28] Junwei Han, King N., Mingjing Li, and Hong-Jiang Zhang, A
Memory Learning Framework for Effective Image Retrieval,
IEEE Trans. on Image Processing, vol. 14, no 4, pp. 511-523,
April 2005.
[29] Marin Ferecatu, Nozha Boujemaa, Michel Crucianu, Semantic
interactive image retrieval combining visual and conceptual
content description,ACM multimedia systems Journal, vol. 13,
No. 5-6, pp. 309-322, 2008.
[30] Anca Loredana Ion, Liana Stanescu, and Dan Burdescu,
Semantic Based Image Retrieval using Relevance Feedback,
international Conference on Computer as a Tool, pp. 303-310,
Warsaw, September 2009.
[31] I Gondra, D.R Heisterkamp, and J. Peng, Improving the
Initial Image Retrieval Set by Inter-query Learning with One
Class SVMs in Proc. of the 3rd International Conference on
Intelligent Systems Design and Applications , Tulsa,
Oklahoma, United States, 2004
[32] Esin Guldogan, Moncef Gabbouj, Dynamic Feature Weights
with Relevance Feedback in Content-Based image Retrieval,
in METU, Northern Cyprus Campus, pp. 56-59. September
2009
[33] Rocchio JJ, Relevance Feedback in Information Retrieval in
the SMART Retrieval system, pp. 313-323, prentice Hall. 1971.
[34] G. Ciocca and R. Schettini, A Relevance Feedback
Mechanism for Content-based Image Retrieval in Inf. Process.
Manage. , vol. 35, no. 5, pp. 605-632, 1999.
[35] I. J Cox, M. L Miller. PicHunter: Bayesian Relevance
Feedback for Image Retrieval in Proc. of the 13th International
Conference on Pattern Recognition, Vienna, Austria, 1996
[36] Michel Crucianu, Marin Ferecatu, and Nozha Boujemua,
Relevance Feedback for image Retrieval: a short survey In
INRIA Recocquent court, B.P 105. Oct. 2004.


Smt. Pushpa B. Patil completed her B. E., and M.Tech. from the Department of Computer Science and Engineering,
Karnataka University Darawad, Visveshwaraya Technological University Belgaum, India in the years 1996, 2006 respectively.
From 1997-2000, she was worked as lecture in Computer Science Department at MBEs Engineering College, Ambajogai,
Maharastra, India. In 2000, she joined as a lecturer in the Department of Computer Science at B. L. D. E s. Institute of
Engineering and Technology, Bijapur, Karnataka, India, where she is presently holding position of Assistant Professor and
doing PhD at S.R.T.M. University, Nanded, Maharastra, India. Her research interests include image processing, pattern
recognition, and relevance feedback in Content Based Image Retrieval. She is a life member of Indian Society for Technical
Education and Institute of Engineers.

Manesh Kokare (S04) was born in Pune, India, in Aug 1972. He received the Diploma in Industrial Electronics Engineering
from Board of Technical Examination, Maharashtra, India, in 1990, and B.E. and M. E. Degree in Electronics from Shri Guru
Gobind Singhji Institute of Engineering and Technology Nanded, Maharashtra, India, in 1993 and 1999 respectively, and
Ph.D. from the Department of Electronics and Electrical Communication Engineering, Indian Institute of Technology,
Kharagpur, India, in 2005. Since June1993 to Oct1995, he worked with Industry. From Oct 1995, he started his carrier in
academics as a lecturer in the Department of Electronics and Telecommunication Engineering at S. G. G. S. Institute of
Engineering and Technology, Nanded, where he is presently holding position of senior lecturer. His research interests include
wavelets, image processing, pattern recognition, and Content Based Image Retrieval.