Research Journal of Recent Sciences ________________________________________________ ISSN 2277 - 2502
Vol. 1(ISC-2011), 415-418 (2012)
Res.J.Recent.Sci.

Mini Review Paper

An Enhanced Approch for Content Based Image Retrieval
Patheja P.S., Waoo Akhilesh A. and Maurya Jay Prakash
BIST, Bhopal, MP, INDIA

Available online at: www.isca.in
(Received 6th October 2011, revised 5th January 2012, accepted 25th January 2012)

Abstract
Image classification is perhaps the most important part of digital image analysis. Retrieval pattern-based learning is the most
effective that aim to establish the relationship between the current and previous query sessions by analyzing image retrieval
patterns. We propose a new feedback based and content based image retrieval system. In this new approach we use neural
network based pattern learning to achieve effective classification and with neural network we use decision tree algorithm to
make less complex mining of images. That approach is more effective and efficient way for image retrieval.
Keywords: pattern-based learning, image retrieval, neural network.

Introduction
Image classification in large data base system is a typical and
complex task as we know. To achieve effective and efficient
method we are propose an approach based on neural network
and decision tree. In this method neural network play an
important roll as pattern recognizer and this patter is
converted as a set of rules by which we can make a fast
image retrieval system when our system is trained. One
algorithm is trained and rules are extracted the classification
is done by using user’s feed back method.
Why we use neural Network: Traditional statistical
classification procedures such as discriminate analysis are
built on the Bayesian decision theory1. In these procedures,
an underlying probability model must be assumed in order to
calculate the posterior probability upon which the
classification decision is made. One major limitation of the
statistical models is that they work well only when the
underlying assumptions are satisfied. The effectiveness of
these methods depends to a large extent on the various
assumptions or conditions under which the models are
developed. Users must have a good knowledge of both data
properties and model capabilities before the models can be
successfully applied. Neural networks have emerged as an
important tool for classification.
The recent vast research activities in neural classification
have established that neural networks are a promising
alternative to various conventional classification methods.
The advantage of neural networks lies in the following
theoretical aspects. First, neural networks are data driven
self-adaptive methods in that they can adjust themselves to
the data without any explicit specification of functional or
distributional form for the underlying model. Second, they
International Science Congress Association

are universal functional approximates in that neural networks
can approximate any function with arbitrary accuracy2. Since
any classification procedure seeks a functional relationship
between the group membership and the attributes of the
object, accurate identification of this underlying function is
doubtlessly important.
Third, neural networks are nonlinear models, which makes
them flexible in modeling real world complex relationships.
Finally, neural networks are able to estimate the posterior
probabilities, which provide the basis for establishing
classification rule and performing statistical analysis3.
Decision tree: Decision trees are used to extract patter from
neural network to use neural network the main key feature of
decision tree are: i. simple to understand, ii. explanation for
the result, iii, Can be combined with other techniques, iv.at
least a solution provide.

Back Round
In digital image classification the conventional statistical
approaches for image classification use only the gray values.
Different advanced techniques in image classification like
Artificial Neural Networks (ANN), Support Vector Machines
(SVM), Fuzzy measures, Genetic Algorithms (GA), and
Genetic Algorithms with Neural Networks are being
developed for image classification.
Techniques of Image Classification: Image classification
plays an important roll for many Studies in science and
different type of environmental applications. There are many
classification algorithms have been developed for
classification of images. In this section we emphasizes on the
analysis and usage of different advanced classification
415

Research Journal of Recent Sciences ____________________________________________________________ ISSN 2277 - 2502
Vol. 1(ISC-2011), 415-418 (2012)
Res.J.Recent.Sci
techniques like Artificial Neural Networks, Support Vector
Machines, Fuzzy Measures, Genetic algorithms and their
combination for digital image classification. Finally the
study depicts the comparative analysis of different
classification techniques with respect to several parameters.
Artificial Neural Network (ANN): Neural Network can
provide suitable solutions for problems, which are generally
characterized by non-linear ties, high dimensionality noisy,
complex, imprecise, and imperfect or error prone sensor data,
and lack of a clearly stated mathematical solution or
algorithm. A key benefit of neural networks is that a model
of the system can be built from the available data. Image
classification using neural networks is done by texture
feature extraction and then applying the back propagation
algorithm.
Textural features, the angular second moment, contrast,
correlation and variance are calculated. After extracting the
textural features the network is trained by standard back
propagation algorithm (BKP).
The back propagation algorithm is implemented in these
steps: i. Initialize weights, ii. Feed input vectors and compute
the weighting sum and then apply sigmoid function, iii.
Calculate error term for each output unit, iv. Calculate the
error term of each of the hidden units, v. Adjust the weights.
Step 2, 3, 4 and 5 are repeated till the error is within
acceptable limits after that it is ready to store for reference
values.
Support Vector Machines: SVM is a supervised learning
process in which data analyze and recognize patterns, used
for classification and regression analysis. A good separation
is achieved by the hyper plane that has the largest distance to
the nearest training data points of any class (so-called
functional margin), since in general the larger the margin the
lower the generalization error of the classifier.
The inductive principle behind SVM is structural risk
minimization (SRM). Risk of a learning machine (R) is
bounded by the sum of the empirical risk estimated from
training samples (Remp) and a confidence interval (ψ): R ≤
Remp+ ψ [8]. The strategy of SRM is to keep the empirical
risk (Remp) fixed and to minimize the confidence interval
(ψ), or to maximize the margin between a separating hyper
plane and closest data points3.
The implementation of SVM required these steps: A
classification task usually involves separating data into
training and testing sets. Each instance in the training set
contains one “target value" (i.e. the class labels) and several
“Attributes" (i.e. the features or observed variables). The
goal of SVM is to produce a model (based on the training
data) which predicts the target values of the test data given

International Science Congress Association

only the test data attributes. Transform data to the format of
an SVM package, Conduct simple scaling on the data,
Consider the RBF kernel K(x; y) =
, Use
cross-validation to and the best parameter C and , Use the
best parameter C and to train the whole training set, Test.
The best parameter might be a selected by the size of data set
but in practice the one obtained from cross-validation is
already suitable for the whole training set4.
Fuzzy Measures: In Fuzzy measures, different relationships
are identified to describe properties of an image. The
members of these property set are fuzzy in their contribution.
The fuzzy measure gives the possibility to describe different
types of stochastic properties in the same form. If the fuzzy
property is more related to a region, then a fuzzy measure is
used. Fuzzy function is used if a stochastic property is to be
described by a particular distribution of gray values. The
fusion of these two stochastic properties is represented as a
fuzzy measure and fuzzy function defines on an area which is
achieved by a fuzzy integral. The result of fuzzy integral is a
new fuzzy measure5. The fuzzy measures are implemented
in these steps: i. Extraction of stochastic properties, ii,
gather Stochastic Information, iii. Apply Fuzzy Functions,
iv. Fusion of Fuzzy Properties by Fuzzy Integrals.
The summation of all combinations of fuzzy measures with
fuzzy functions makes sure, that all possible properties in all
combinations, which should be considered, are used. In such
a way an image is obtained, where the (grey) values
represent a measure for the membership to the texture. In this
way different elementary stochastic properties are combined
in many ways for the extraction of relevant information. In
order to achieve this different approaches have to be applied
for the elimination of the elementary stochastic properties
within an image.
Genetic Algorithms: The features like texture or the average
value of nearby pixels are necessary to get good spectral
information. The different kinds of spatial content
information could also be added into the pixel feature vector
as additional feature dimensions. So there are a large number
of choices for additional feature vectors that could make
classification better than just having the raw spectral values
as feature vectors.
Data

Feature

Image processing

Supervised learning

GA

Output

Figure-1
Process of genetic algorithm process

416

Research Journal of Recent Sciences ____________________________________________________________ ISSN 2277 - 2502
Vol. 1(ISC-2011), 415-418 (2012)
Res.J.Recent.Sci
The genetic programming system based on a linear
chromosome that manipulates image processing programs
that take the raw pixel data planes and transform them into a
set of feature planes. This set of feature planes is in effect
just a multi-spectral image of the same width and height as
the input image, but perhaps having a different number of
planes, and derived from the original image via a certain
sequence of image processing operations. The system then
applies a conventional supervised classification algorithm to
the feature planes to produce a final output image plane,
which specifies for each pixel in the image, whether that
feature is there or not. Figure illustrates this hybrid scheme.
In this structure finally raw data planes are transformed into a
set of feature planes by an image processing program that is
evolved by genetic algorithm6.
Genetic Algorithms and the Neural Network: The
technical design of the evolutionary strategy of connection
weights training can be described as: i. Decode each
individual (genotype) in the current generation into a set of
connection weights, ii. Decode each individual (genotype) in
the current generation into a set of connection weights,
iii.Evaluate each set of the connection weights by
constructing the corresponding neural network structure and
computing its total mean square error, iv. Select parents for
reproduction based on their fitness, v. The population of
current generation is mapped onto a roulette wheel, vi. Apply
search operators in conjunction with the crossover and /or
mutation operator.
Comparative Analysis: The image classification techniques
like artificial neural networks, support vector machines,
fuzzy logic, genetic algorithms and their combination are
analyzed and compared with respect to several parameters.
Artificial neural networks have the advantages mainly of
more tolerance to noise inputs and representation of Boolean
function apart from others. But too many attributes may
result in over fitting. In support vector machines over fitting

is unlikely to occur. The computational complexity and
complexity of decision rule are reduced in SVM. Fuzzy
measures have the benefit of identification of various
stochastic relationships to describe the properties of the
image. But priori knowledge is very important to get good
results.
Genetic algorithms are primarily used in optimization and
always have a good solution. But the computation of scoring
function is non trivial. The artificial neural networks and
support vector machines follows non-parametric approach
whereas fuzzy measures use stochastic properties for image
classification. The selection of non-linear boundary is
efficient when the data have only few input variables in ANN
and vice versa in SVM. In fuzzy logic it depends on priori
knowledge.
Where as in genetic algorithms it depends on the direction of
decision. The training speed in the neural networks depends
on network structure, momentum rate, learning rate and
converging criteria. In SVM it depends on training data size
and class reparability. Fuzzy logic incorporates the training
speed depending on the isolation of the relevant information
by iterative application of the fuzzy integral. The training
speed could be improved by refining irrelevant and noisy
genes in genetic algorithms. Along with these the parameters
accuracy and general performance are tabulated in table4.
Praposed Model: In this paper we are discuss various
methods and techniques by which image classification
performed in effective and efficient manner. Here we are
going to propose a new way for classification for image
retrieval system. In this model we use neural network,
decision tree and user feedback for classification of
image.We proposed a new approach for Content based image
retrieval system feedback based image classifier in efficient
manner.

Table-4
Comparative analysis of image classification techniques
Parameter

Artificial neural network

Support vector machine

Fuzzy logic

Genetic algorithm

Type of approach

Non parametric

Non parametric with binary
classifier

Stochastic

Large time series data

Non linear decision
boundaries
Training speed

Efficient when the data
Depends on prior
Efficient when the data have
Depends on the direction
have only few input
knowledge for
more input variables
of decisions
variables
decision boundaries
Network structure,
Training data size, kernel Iterative application Referring irrelevant and
momentum rate ,learning
parameter
of the fuzzy integral.
noise genes.
rate, converging criteria

Accuracy

Depends on number of
input classes.

Depends on selection of
optimal hyper plane.

Selection of cutting
threshold

Selection of genes.

General
performance

Network structure

Kernel parameter

Fused fuzzy integral.

Feature selection.

International Science Congress Association

417

Research Journal of Recent Sciences ____________________________________________________________ ISSN 2277 - 2502
Vol. 1(ISC-2011), 415-418 (2012)
Res.J.Recent.Sci
We have prepared a database consisting a number of images.
Here user required to input the test image or query image and
selected object images. now Features like color and texture
are then extracted from test (query) image and object images.
Similarities distances are measured and calculating linear
Coefficient of Correlation between tests (query) and object
images. On the basis of coefficient of correlation Training set
is formed into two categories (relevant and irrelevant). Then
using those labeling the neural network based classifier is
trained7.

Conclusion
Here, the above approach is just a proposal about a new
system. In future we design a tool for the above given
method for image classification and provide compression for
that system.

References
1.

These feedback rounds go on iteratively and each time a
refinement of the images shown to the user is done. Finally
those images which are judged by the classifier as positive
are taken. The classifier results are produced in a decision
tree algorithm to extract rules from it. Now the query image
is plotted in the positive set.

Gupta Neetesh, Singh R.K. and Dubey P.K., A New
Approach for CBIR Feedback based image classifier,
International Journal of Computer Applications (0975
– 8887) 14(4), (2011)

2.

Gilbert Adam D., Chang Ran, and Xiaojun Qi, A
retrieval pattern-based inter-query learning approach
for content-based image retrieval, Proceedings of 2010
IEEE 17th International Conference on Image
Processing, (2010)

Those images which are nearest to the query image based on
these rules set similarity ranking are taken and shown to the
user. We are going to follow these steps for our proposed
technique.

3.

Chih-Wei Hsu, Chih-Chung Chang, and Chih-Jen Lin
Department of Computer Science National Taiwan
University, Taipei 106, A Practical Guide to Support
Vector Classication, Taiwan, Initial version, 2003,
(2010)

4.

Ajitha Gladis K.P. and Ramar K., A Novel Method for
Content Based Image Retrieval Using the
Approximation of Statistical Features, Morphological
Features and BPN Network, IEEE computer society
ICCIMA, 148, 179-184 (2007)

5.

Liu P., Jia K., Wang Z. and Lv Z., A New and
Effective Image Retrieval Method Based on Combined
Features, Proc. IEEE Int. Conf. on Image and
Graphics, I, 786-790 (2007)

Relevance Feedback: Mark all Images as Relevant Images
as well as Irrelevant Images and Forms set Rr and Ri. (Where
Rr ε Tr and Ri ε Tr) and Tr is Training Set.

6.

Training and Rule Extraction: The set Rr is modified by
including the query image selected by user in it. Calculate
feature vector of Rr and Ri. Feature vectors of Fr forms the
positive set and feature vectors of Fi forms the negative set
of data points for training the classifier.

Tienwei Tsai, Te-Wei Chiang and Yo-Ping Huan,
Image Retrieval Approach Using Distance, Threshold
Pruning, IEEE Trans. On Image Processing, 12, 241249 (2007)

7.

Guoqiang Peter Zhang, Neural Networks for
Classification, A Survey IEEE transactions on systems,
man, and cybernetics—part C, applications and
reviews, 30(4), (2000)
Banzhaf W., Nordin P., Keller R.E. and Francone F.D.,
Genetic Programming: An Introduction, Morgan
Kaufmann, San Francisco, CA, (1998)

Data Preprocessing: Initialize with various object images,
Create a Block Matrices, Calculate Mean μ of Block
Matrices and Concatenate.
Feature Extraction: Convert Block Matrices “f & g” RGB
from space to HSV from space, where f and g represent the
average values of vectors, Extract feature vector Vj from
HSV space. And combined all color and texture features.
Similarity Computation: Calculate Euclidean Distance then
get Euclidean, Repeat above procedure for n object images
now we have “N” object image and its Euclidean distance
Matrices.

These sets are then given as input to classifier. The feature
set of Tr is calculated and then fed to classifier for
classification so that the separate data points in training set as
positive or negative. After that classified data we apply
decision tree for explanation of rules. This Process is iterated
many times. According to user input.
Result Processing: Collect the set FTr from the last iteration
of training. Now, say user wants g numbers of images from
the database which are most relevant to the query image

International Science Congress Association

8.

9.

Richard M.D. and Lippmann R., Neural network
classifiers estimate Bayesian a posteriori probabilities,
Neural Comput., 3, 461–483, (1991)

10.

Duda P.O. and Hart P.E., Pattern Classification and
Scene Analysis, New York, Wiley (1973)

418

