WEB IMAGE RETRIEVAL USING CLUSTERING APPROACHES

David Bracewell, et al. (Eds): AIAA 2011,CS & IT 03, pp. 27–36 , 2011.
© CS & IT-CSCP 2011 DOI : 10.5121/csit.2011.1304
WEB IMAGE RETRIEVAL USING CLUSTERING
APPROACHES
Umesh K K1
and Suresha2
1
Department of Information Science and Engineering,
Sri Jayachamarajendra College of Engineering, Mysore, India
umeshkatte@gamil.com
2
Department of Studies in Computer Science,
University of Mysore, Manasagangothri, Mysore, India
sureshabm@yahoo.co.in
ABSTRACT
Image retrieval system is an active area to propose a new approach to retrieve images from the
large image database. In this concerned, we proposed an algorithm to represent images using
divisive based and partitioned based clustering approaches. The HSV color component and
Haar wavelet transform is used to extract image features. These features are taken to segment
an image to obtain objects. For segmenting an image, we used modified k-means clustering
algorithm to group similar pixel together into K groups with cluster centers. To modify K-
means, we proposed a divisive based clustering algorithm to determine the number of cluster
and get back with number of cluster to k-means to obtain significant object groups. In addition,
we also discussed the similarity distance measure using threshold value and object uniqueness
to quantify the results.
KEYWORDS
Feature Extraction, Pixel Segmentation, Modified k-means Clustering, Similarity Distance
Measure, Object Clustering, Object Uniqueness.
1. INTRODUCTION
Due to rapid exponential growth of image data and its related technologies, many users gets
motivated to browse and search images from the image database. So that it is very essential to
develop innovative tools for managing and searching the images. Most of the end users use an
existing search engine to browse the information to looking the similar documents. These
documents has provides similar kind of information to the end user. This is true for those who are
looking the text related documents can be get from the web by sending query in the form of set of
keywords but it is not true for the intuition of web image browsers to retrieve similar images, also
it is not possible to achieve to retrieve based on the contents of the image.
We have many search engines on the Web to retrieve images and they provide an interface to
receive the query in the form of set of keyword [1], [2], [3]. Most of retrieved images are similar
but it is not relevant to the user query in most of the cases. So that it needs to be process the
content of an image to represent with some meaning. There are few challenges in the web and it
can be solved by using the existing techniques of Content Based Image Retrieval system. We are

28 Computer Science & Information Technology (CS & IT)
motivated to utilize the technique of CBIR to implement our proposed model to retrieve Web
images.
In recent years, many researchers have been involved in the area of Content-based image retrieval
(CBIR) from unannotated image databases [4]. Now a days, enormously many people uses a
Digital image and video libraries as a visual information. The technologies needed for retrieving,
managing and browsing this growing amount of information are still open issues; however, quite
still there is an open challenge for the research community. Many projects have been started in
recent years to research and develop efficient systems for content- based image retrieval. The
best-known CBIR system is probably Query by Image Content (QBIC) [5] developed at the IBM
Almaden Research Center. Other noteworthy systems include MIT's Photobook [6] and its more
recent version, FourEyes [7], the search engine family of VisualSEEk, WebSEEk, and MetaSEEk
[8],[9],[10], which all are developed at Columbia University. The Virage [11] is a commercial
content-based search engine developed at Virage Technologies.
Our proposed model uses a cluster based approach to compute the centroids for every image. This
is used for image representation. The main advantage of this algorithm is very easy and takes less
time to preserve the cluster of pixels. Basically K-Means is unsupervised and iterative clustering
algorithm, which finds the optimal set of cluster. The retrieval method is based on the similarity
measure is discussed in detail in the section 2.4.
The organization of the paper is as follows: In section 2, the Methodology of Extracting Image
Features, Pixel Segmentation, Object Clustering, Similarity Distance Computation and
Implementation are discussed. In Section 3, we discussed the conclusion of the paper.
2. METHODOLOGY
The block diagram of image retrieval system is shown in figure 1. It consists of two phases: off-
line and On-line.
In Off-line, feature extraction model extracts color and texture features. These features are
inputted to the clustering module to find the centroids of every image, and then stored in the
centroids database.
Input Image Set of similar
Images
Off-line
Query Image
On-line
Figure1: The Block Diagram of Image Retrieval system.
In on-line, Image searching is actually processed in the retrieval phase. It accepts query image
and compares with image features using threshold and object uniqueness to measure the
similarity to retrieve the images from the image database.
In brief, in this paper we discussed about three important task of the proposed system.
1. Image pre-processing module, it helps to obtain image features. These features are used
to segment to get objects/regions by grouping pixels with similar descriptions.
2. Clustered the objects to tune up the speed of the retrieval system.
Feature
Extraction
Divisive
Clustering
K-means
Clusterin
Centroids
Database
Similarit
y
Measure

Computer Science & Information Technology (CS & IT) 29
3. Employed similarity distance measure to determine the distance between the image query
and image features database.
2.1. Feature Extraction
In this section, we discuss how to extract the image features of color and textures. According to
Li and Wang [12] and Zhang [13], image consists of large number of pixels. The values of pixels
are redundant; it can be reduced in the pre-processing stage of an image. In the pre-processing
stage, to achieve good result the image is portioned into 4 by 4 blocks; each block is replaced by
the average of 4 by 4 pixel values to compromise between texture granularity, segmentation
coarseness, and computation time. The HSV color space is widely used for extracting color
features due to its ability to transform from RGB to HSV and vice versa. Since HSV color space
has an advantage to produce a collection of colors that is also compact and complete. These
features are denoted as {f1, f2, and f3}.
To obtain other three texture features, we applied Haar wavelet transform to the L component of
an image. The Haar wavelet is discontinuous and resembles a step function. It represents the
energy in high frequency bands of the Haar wavelet transform [14]. After one-level wavelet
transform, a 4 by 4 block is decomposed into four frequency bands, each band containing a 2 by 2
matrix of coefficients, suppose the coefficients in the HL band are {ck+i,ck,l+1,ck+1,l,ck+1,l+1}.
Then the feature of the block in HL band is computed using equation 2.1:
f = ቂ
ଵ
ସ
∑ ∑ c୩ା୧,୪ା୨
ଶଵ
୨ୀ଴
ଵ
୧ୀ଴ ቃ
భ
మ
2.1
The other two features are computed similarly in the LH and HH bands. These three features of
the block are denoted as {f4, f5, f6}.
Let, Image I represented as ‫ܫ‬ = ሼܴଵ, ܴଶ, … ܴ௡ ሽ, ‫ݔ‬௜ ∈ ܴௗ
be a region based feature vector set.
These regions feature have extracted by using K-means algorithm. To find the signature of ܰ
weight vectors, that is ܿ = ሼܿଵ, ܿଶ, … , ܿ௡ሽ as representatives of the feature vector of I= {f1, f2, f3, f4,
f5, and f6}.
2.2. Pixel Segmentation
In section 2.1, we discussed the procedure to get the required features. These six features are
taking into consideration for segment an image into objects. For segmenting an image, we
modified K-Means clustering algorithm to group similar pixel together into k groups with cluster
centers. This is quite efficient and minimizes the sum square error is defined in the equation 2.2.
Dሺkሻ = ∑ ݉݅݊ଵஸ୨ஸ୩൫x୧ − xത୨൯
ଶ୪
୧ୀଵ 2.2
The advantage of K-Means algorithm is simple and quite efficient. It works well when clusters
are not well separated from each other [15]. This could be happen in web images. But the
disadvantage of K-means is expected to receive the value of k from the user; each value of k is
represented by their centroids. To certain extent for selecting the value of k requires domain
knowledge. If the value of k is selected randomly, it gives different distinct total Sum Squared
Errors. This is not a good idea to choose the value of k randomly to make a good cluster. This is
the issue mainly we focused in this paper to choose the value of k of K-Means dynamically to
determined the number of k values instead of randomly chosen.
This is the reason we used divisive based algorithm to obtain initial cluster centers. In this
approach, first merges pair of the closest pixels to form cluster. These clusters are then merged to

generate a new bigger cluster and finally form a single hierarchical cluster that covers the whole
image.
To calculate the distance between the new cluster and remaining clusters, we use the average
linkage method. Let ܿ௡ be a new cluster produced by merging of clusters ܿ௜ and ܿ௝. Let ܿ௞ be a
remaining cluster. The distances between the new cluster and the remaining clusters are computed
using equation 2.3.
Distሺc୬, c୩ሻ =
|ୡ౟|
|ୡ౟|ାหୡౠห
Distሺc୬, c୩ሻ +
|ୡ౟|
|ୡ౟|ାหୡౠห
Dist൫c୨, c୩൯ 2.3
We used person correlation coefficient function to find distance between each pixel to obtained
better cluster than Euclidean distance during image segmentation. Let ‫ݔ‬௜ be the features of pixel1
and ‫ݕ‬௜ be pixel2, where i=1...6. To find the distance between the pixel x and y is computed using
the person correlation coefficient. Since the number of features per pixel is 6. We set different
weight between color (Wc) and texture (Wt) to get better clustering result.
2.3. Object Clustering and Similarity Distance Computation
In this section, we discussed about object clustering and similarity distance computation. The
scope of this discussion is to increase the speed of image retrieval system. We cluster similar
objects to compare the query image to all images in the database. This takes more time and it
consumes more memory, when the database is large. To reduce the computational cost of image
retrieval system, we perform hierarchical clustering to the entire objects. Each object is identified
by six features, which are the average features of all the member pixels.
The query image goes through the same image segmentation algorithm to obtain objects. These
objects are compared to the cluster centers in the database and similarity is calculated using L2
distance. The objects in the database that has a minimum distance will be returned to perform
global image distance computation between query image and database image. The objects in the
database that has a minimum distance will be returned to perform global image distance
computation between query image and database image.
Figure2.2: Overall similarity distance between image query and image database.
Fig 2.2 shows that, image1 is the query image and image2 is the returned image from the image
database. To determine the overall distance between the query image and the returned image, we
should match all the objects from the two images. From these two images, we measured the
distance between the first object from image1 (Oଵଵ) and the first object from image2 (Oଶଶ), is
defined as:
Oଵଵ,ଶଵ = ∑ ሺfଵଵ୧ − fଶଵ୧ሻଶ଺
୧ୀଵ 2.4
where fi is the color and texture from each object.

The similarity distance computation between image1 and image2 can be defined as:
w୮ୀ ஑౦ஒ౦ஓ౦
2.5
Where α୮ = 1 when the two region being match is the closest distance
α୮ = 0 when the two region being matched is not the closest distance
β୮ = (#of pixels for object p) / (#of pixel in the whole image)
γ୮ = Object uniqueness of the object based on the appearance of the object
We determine the uniqueness of the object is based on the visual appearance of the object in the
object cluster. When an object appears frequently in the database, it can be consider less unique
and vice versa. To get the value of γ୮ , first we carried out hierarchical clustering to group similar
objects in the same group. Next, we determine the uniqueness of object p (γ୮ ) based on the total
number of objects that belong to the cluster compared to the total number of objects in the
database. The similarity distance between ‫ܦ‬ሺ‫,1݉ܫ‬ ‫2݉ܫ‬ሻ and ‫ܦ‬ሺ‫,2݉ܫ‬ ‫1݉ܫ‬ሻ is not symmetric,
therefore in order to make it symmetric, we take the average between them. The overall distance
is defined as:
Overall DሺIm1, Im2ሻ =
ୈሺ୍୫ଵ,୍୫ଶሻା ୈሺ୍୫ଶ,୍୫ଵሻ
ଶ
2.6
By using this equation, it is straightforward to determine ‫ܦ‬ሺ‫,ܺ݉ܫ‬ ‫ܻ݉ܫ‬ሻ for every images in the
database. This is most well balanced technique in similarity measure between regional and global
matching. This cluster of object groups is necessary for two reasons: for faster image retrieval and
for determining the value of object uniqueness is discussed in the subsection 2.4.2
2.4. Implementation
To retrieve relevant images, user sends a query in the form of image to the image retrieval system
to find out the set of similar images. Whenever it receives the query, it sends to the feature
extraction module to compute the signature of the image. The signature of the image is compared
with features vectors stored in the database using the similarity measure. The similar images are
retrieved and then displayed the corresponding thumbnail images at the user place. These
thumbnail images are available in the database for the user perspective point of view in the form
of JPEG format.
2.4.1. Performance
Currently, we have conducted an experiment with two image databases of different sizes. The
first database contains 1,000 images with size of 384 X256 resolution. The second database
contains 10,000 low resolution web-crawled miscellaneous image database with the size of
128 X 85 resolution. Most of the images are color photographs in JPEG format and used in
WBIIS download from the image collection residing at http://wang.ist.psu.edu/docs/related.shtml.
There are 10 categories, around 100 images per category in the first database and 100 categories,
around 100 images per categories in the second database.
We used a Pentium(R) 4 CPU 3.00 GHz PC using the Windows-XP operating system. It takes
approximately 3 hours to compute the feature vectors. On average, 0.003 seconds is needed to
compute the features of all blocks of an image. The feature clustering process is performed by
only once for our database. The clustering procedure we have discussed with detail in section 2.3.

We have conducted an experiment to measure the recall and precision of the system based on the
images which are presented in the database as a query image. Similarly, images which are not
present in the image database as a query image. We obtain better precision for the given query
image shown in figure 2.3, 2.4, 2.5 and 2.6. Due to the limitations of space, we shows only two
rows of images with the top 10 matches of sample query.
2.4.2. Experiments
In pre-processing stage, we used MATLAB image processing tools to extract the pixel features
on each image.
2.4.2.1. Experiments #1-Similarity Distance Measure.
We obtained objects during pixel clustering. To find a suitable number of objects per image, we
need to set the similarity distance measure during divisive clustering. To do this, we randomly
selected 10 images from each image class which corresponds to 100 images from the database.
We ran this experiment for one time. From each image, we set the similarity measure at 0.6, 0.7
and 0.8. The number of clusters based on these measures was recorded and the average number of
clusters for each similarity measure was computed is shown in Table 2.1.
Table 2.1 Correlation between similarity measures and average number of objects
similarity
measure
Average number of
objects
0.6 3.3
0.7 6.4
0.8 9.2
We concluded to set a similarity measure equal to 0.7 because at similarity measure 0.7 the
average number of clusters for each image is 6.3 which is a good number of objects. We argue
that 6.3 is a good number of objects for this particular database because when we randomly select
100 images and visually counted the number of objects per image, the average is 6.3. To see the
performance of experiment #1 we randomly select 2 images from different class, namely horse
and flower. Each query returns the top 10 images from database. The two query retrievals are
shown in Figure 2.2 and 2.3
Figure 2.2: Best top 10 matches of a sample horse image query using similarity distance.

Figure 2.3: Best top 9 matches of a sample flower image query using similarity distance.
2.4.2.2. Experiments #2 – Object Uniqueness
The idea of object uniqueness is to give more weight to objects that are more distinctive
compared to other objects that are more common. To do this, we clustered all objects in database.
Objects that belong to larger clusters will get less uniqueness weight compare to objects that
belong to smaller clusters.
The selection of the number of clusters in object clustering affects the result of accuracy during
image retrieval. The parameter of object uniqueness is obtained from the object clustering. Since
object uniqueness is one of the variables to compute similarity distance computation, therefore, to
select the number of cluster that results in the best performance. Therefore, during the experiment
we choose four different number of cluster during object clustering: 1, 10, 25, and 35. All 100
images from each image class are selected for query, and each query returns the result of top 10
from the database. For each query we examine the precision of the returned images on their
relevance of semantic meaning. Since similar semantic meaning belong to the same class, and
each class has the same number of images, therefore it is straightforward to calculate the
precision and recall. The precision and recall using different numbers of object clusters are then
compared is shown in table 2.1.
Table 2.1: Average precision/recall comparison of top 1, 10, 25, and 35 closest distances.
Image
Id
object
cluster(1)
object
cluster(10)
object
cluster(25)
object
cluster(35)
1 0.22 0.27 0.31 0.3
2 0.28 0.31 0.35 0.34
3 0.29 0.35 0.42 0.423
4 0.75 0.61 0.76 0.75
5 0.89 0.96 1.02 0.97
6 0.6 0.59 0.78 0.76
7 0.31 0.28 0.32 0.32
8 0.41 0.4 0.42 0.41
9 0.34 0.31 0.36 0.35
10 0.41 0.39 0.44 0.42
From Figure 2.4 we examined that the precision and recall is higher in almost all image class
when the number of object cluster is set to 25. The outcome of 35 object clusters is better
compared with 1 object cluster, while the result for 10 object cluster is worse compare to the rest
of the cluster number. From this result we conclude that the parameter of object uniqueness in

similarity distance computation increase the performance of image retrieval. However, to get the
best result, we need to find the accurate number of cluster during object clustering.
Figure 2.4 Correlation between similarity measures and average number of objects
To see the performance of Experiment #2, we randomly select 2 images from different class,
namely horse and flower. Each query returns the top 10 images from database. The two query
retrievals are shown in Figure 2.5 and 2.6
Figure 2.5: Best top 10 matches of a sample horse image query.
Figure 2.6: Best top 10 matches of a sample flower image query.

3. CONCLUSIONS
We developed image retrieval system using clustering approaches. Our proposed model uses a
modified clustering algorithm to improve image segmentation and discussed new similarity
distance measure where object is considered during computation. This increases the speed of
retrieval system to achieve better precision and recall. We have conducted experiments to
evaluate the performance of the system by similarity measure using a threshold value and object
uniqueness. We observed in experiment #2, the object uniqueness measures gives better results
than experiment #1, because it gives more weight to the object that are more distinctive compared
to other objects. The performance of the system is evaluated based on similarity measures using a
threshold value and object uniqueness. Our proposed model gives better result in some of the
classes and performs worse when the image is complex and the objects have smooth edges.
During the implementation, we also proved that by considering object uniqueness during
similarity distance computation improve the accuracy during retrieval. This work can be extended
to classify the images using an advanced classifier to achieve better performance for a particular
class.
REFERENCES
[1] http://images.google.com/
[2] http://www.ditto.com/
[3] http://www.altavista.com/
[4] Rui, Y., Huang, T.S., Chang, S.-F., 1999. Image retrieval: current techniques, promising directions,
and open issues. Journal of Visual Communication and Image Representation 10 (1), 39-62.
[5] Flickner, M., Sawhney, H., Niblack, W., et al., 1995. Query by image and video content: the QBIC
system. IEEE Computer September, 23-31.
[6] Pentland, A., Picard, R.W., Sclarof, S., 1994. Photobook: tools for content-based manipulation of
image databases. In: Storage and Retrieval for Image and Video Databases II. In: SPIE Proceedings
Series, Vol. 2185. San Jose, CA, USA.
[7] Minka, T.P., 1996. An image database browser that learns from user interaction. Master's thesis,
M.I.T., Cambridge, MA.
[8] Michael J. Swain., Charles Frankel., and Vassilis Athitsos, “WebSeer: An Image Search Engine for the
World Wide Web”, Technical Report 96-14, 1997.
[9] J. R. Smith, “Integrated Spatial and Feature Image Systems: Retrieval, Compression and Analysis”.
PhD thesis, Graduate School of Arts and Sciences, Columbia University, February 1997.
[10] S. Sclaroff., L. Taycher., and M. La Cascia. “Imagerover: A content-based image browser for the
world wide Web”. In Proceedings IEEE Workshop on Content-based Access of Image and Video
Libraries, June ’97, 1997.
[11] Bach, J.R., Fuller, C., Gupta, A., et al., 1996. The Virage image search engine: an open framework for
image management. In: Sethi, I.K., Jain, R.J. (Eds.), Storage and Retrieval for Image and Video
Databases IV. In: SPIE Proceedings Series, Vol. 2670. San Jose, CA, USA.
[12] Li, J., Wang, J. Z. and Wiederhold, G., (2000), “Integrated Region Matching for Image Retrieval,”
ACM Multimedia, p. 147-156.
[13] Zhang, R. and Zhang, Z., (2002), “A Clustering Based Approach to Efficient Image Retrieval”,
Proceedings of the 14th IEEE International Conference on Tools with Artificial Intelligence, pp. 339.
[14] Daubechies, I., (1992), “Ten Lectures on Wavelets,” Capital City Press, Montpelier,Vermont.
[15] Guha, S., Rastogi, R., and Shim, K., (1998), “CURE: An Efficient Clustering Algorithm for Large
Databases” Proc. of ACM SIGMOD International Conference on Management of Data, pp.73-84.

Authors
Umesh K K is a faculty member with
the Department of Information
Science and Engineering, Sri
Jayachamarajendra College of
Engineering, Mysore. He is also a
Research Scholar, Department of
Studies of Mysore and M.Tech
degree in 2003 from Visvesvaraya
Technological University, Karnataka.
He is working towards doctoral
work under the supervision of Dr.
Suresha.
Dr.Suresha is presently working as
Associate Professor, Department of
Studies in Computer Science,
University of Mysore. He received his
B.Sc. degree in 1987 from University
of Mysore, M.Sc degree in 1989 from
University of Mysore and M. Phil
degree in 1991 from DAVV, Indore.
He received M.Tech degree from
Indian Institute of Technology,
Kharagpur in 1996 and Ph. D. from
prestigious Indian Institute of Science,
Bangalore in 2007. He was awarded
second prize in IRISS-2002
competition, which is an all India level
research student competition called
“Inter Research Institute Student
Seminar”. He has a teaching
experience of 21 years at post graduate
level, has 12 publications to his credit
and currently supervising five research
scholars towards their doctoral work.

WEB IMAGE RETRIEVAL USING CLUSTERING APPROACHES

More Related Content

What's hot (19)

Similar to WEB IMAGE RETRIEVAL USING CLUSTERING APPROACHES (20)

More from cscpconf (20)

Recently uploaded (20)

WEB IMAGE RETRIEVAL USING CLUSTERING APPROACHES