Ijcs 2016 0303012 PDF

International Journal of Computer Systems (ISSN: 2394-1065), Volume 03 Issue 03, March, 2016
Available at http://www.ijcsonline.com/
An Image Retrieval Approach based on Colored Regions Features and a

Clustering Technique
Fahd A. Alqasemi, Amira Abdelwahab and Hatem Abdelkader
,,
Information Systems Department,

Menoufia University, Egypt
Abstract
Image retrieval has important trends in recent decades, due to image means, semantics, and implications. Since
searching in images and retrieving target image has achieved based on text annotated the image, this type of image
retrieval had faced many problems and errors, which led to the emerging approaches of detecting similarity properties
between images. These approaches has used image content rather than image annotations, which called content-based
image retrieval. Content-based image retrieval was achieved using image semantic features, which are shapes, colors,
and texture. This paper, presents an approach used clustering techniques of these features to measure the similarity and
giving results. Clustering is done by k-means algorithm. Using image correlation based on fixed referenced image was
one of three global features used here, these global features had given good results, however, they do not give any
information about specific regions, objects, and colors in the image, therefore we has used shape signature with its own
pre-classified color that represent specific features in the image, and draw with global features the whole idea about
images, wither these in image test set or target image to be look for similarities in the test set . Test set that experiments
had carried on, has divided into six categories, which made the diversity of results and give clear idea about the effect of
this approach in different types of images.
Keywords: CBIR, Global features, Local features, Shape signature, Image semantic level, K-means.
I.
INTRODUCTION
Content-based image retrieval (CBIR) is used computer

vision methods for the reason of image retrieval problem
solving. CBIR helped in the challenges of searching for
images in big databases. CBIR was not used traditional
image searching approaches, which exploit text annotated
the image to detect the searching results.
CBIR used pixels information to gain an idea that is
high level information (semantic level) about image
contents, some researchers has used human assistant as a
feedback to improve the direction of the search, but in this
paper there are no feedback no text annotation to images,
these classification of image retrieval types is depicted in
Figure 1.
The related researches used image features to gain
information about the images, these features are about
texture, shape or color, here the texture features has used
with two additional ideas, i.e. using image correlation
based on fixed referenced image, and using shape signature
with its pre-classified color.
Significant CBIR problem is in the semantic level of
image data, i.e. values of pixels stored in bytes, this mean
the pixels are hard to interpret in machine comparing with
human observers [1].In contrast with text where the textual
data are closed to human understanding. They tried to solve
this problem by using semantic based image retrieval,
which is relied on CBIR.
Semantic gap (Figure 2) is the problematic distance
between the high-level image content, which human
interprets easily, and the low-level features, which has used

for describing images[2].There are several experiments to
bridge the semantic gap, in what they called semantic based
image retrieval (SBIR), which exploit the techniques of
CBIR in a practical semantic manner [3].
In this paper we proposed an approach that used two
minor techniques, the aim is to exploit each technique
solely, then grouping the power of them to propose a total
value. At first we applied image global features technique,
then image local technique, and calculate the results of
each technique solely, after that we group them in one final
results. These experiments were not applied to one data set
(images set), however, we used five categorizes. All that
indicated two important issues, the significant of using any
type of features solely, and the importance of supporting
the final results with these two types of features.
In additional to this, we has used clustering technique
called k-means, for discriminating and grouping varies
results in all experiments.
All results has presented via graphs whether these
related to categories results, or those related to performance
measurement.
II.
RELATED WORKS
Traditional Image retrieval could be done using

annotated images, this annotation has made manually, as in
[4], however, recent papers used content based image
features simultaneously with annotated images.
In [5],Singh et al, had proposed a new query-byexample technique that has exploited multiple shape,
253 | International Journal of Computer Systems, ISSN-(2394-1065), Vol. 03, Issue 03, March, 2016
Fahd A. Alqasemi et al
An Image Retrieval Approach based on Colored Regions Features and a Clustering Technique
texture and color features. It has used images texture

property, color histograms and shape features. The
matching individually then jointly has made.
In [4], Gangan et al, had presented an approach that
called automatic image annotation (AIA) using MPEG-7
features. Papers Application domain was fruit pictures. In
AIA, the image contents has labeled automatically using a
pre-defined set of concepts, which has used for
representing the image implicit meanings.
In [6], Ja-hwunsu et al, had proposed a pattern-based
conceptual image retrieval method (named PBCIR)for gain
visual patterns from visual features into visual pattern
matching has used by relevant images and concepts, which
can be extract to fulfill of semantic image retrieval goal.
Image
Test Set
Features
Extraction
Target
Image
Features
Extraction
Matching
and
Resulting
Results
Figure 3. Semantic-based CBIR project framework
IV.
BACKGROUND
The features that extracted are global features and then

specific (local) feature. Global features also called image
texture features.
The texture features we used are:
A.
Image standard deviation

The Standard Deviation is a measure of how spread out
numbers are. The numbers used here are the pixels of the
image, mathematically; standard deviation is the square
root of the Variance.
Figure 1.
Image retrieval approaches
Standard deviation is used usually to detect and distinct

between the high frequency component and low frequency
component in the image, which help to find the similarity
between images [8].
In [3], Kalantidis et al, had presented a region-based

approach due to image retrieval semantically. Image
segmentation and the popular Bag-of-Words model has
combined.
In [7], color histogram and DBC co-occurrence matrix
for CBIR are utilized. The HSV color histogram
alsoutilized to form the feature database. Their results show
good improvement as their evaluation measures as
compared to other transform domain features.
III.
(1)
B. Image entropy
Entropy is a statistical measure of randomness that can
be used to characterize the texture of the input image [8].
Entropy has defined as
SYSTEM FRAMEWORK
(2)
CBIR framework has consisted of three general steps

[2]:
Image test set features extraction.
Target image features extraction.
Matching and resulting.
There are no user feedback, that used in the process of
features, which gained and results that has gotten, so this
system has made CBIR process semantically ,i.e. without
any feedback or user assistant or any image annotation [2],
this framework has illustrated in Figure 3.
C.
Image correlation with a referenced image.

Correlation refers to any of a broad class of statistical
relationships involving dependence, n statistics and
probability theory, correlation means how closely related
two sets of data are [8]:
(3)
A referenced image has chosen to measure a correlation
between it and each image in image test set.
Specific features are:
D.
Figure 2. Semantic gap in image retrieval
Shape invariant moment

A statistical method to get a signature of a shape from
an image, its logarithm produces a vector of seven values
that suppose do not change even the shape resized or
rotated.
Image
Test Set
Texture: std, entropy, correlation

Extraction
Color & region features
Clustering region and color features
Clustering general features
Re-classify Target image texture

information
Re-classify Target image region and

color information
Matching and resulting

(i)
Matching and resulting

(ii)
Target
Image
Retrieval
results
Figure 4. Diagram illustrates the process of proposed system.
E. Shape color class

We use a way to cluster the colors of the images shapes
by computing the average of the shape pixels, and cluster
all shapes to determine the color class of the each shape.
F. XYZ color standard
XYZ is a color model that was derived from a series of
experiments, this model defines Y as luminance. Z is quasiequal to blue stimulation, and X is a linear combination of
eye cone response curves chosen to be nonnegative.
Defining Y as luminance has the useful result that for any
given Y value, the XZ plane will contain all possible
chromaticities at that luminance [8].
G.
K-means algorithm for information clustering

K-means is an algorithm that is popular for cluster
analysis in data mining [9]. It has used here for
discriminating and ordering features extracted from
image set and then from target image.
k-means clustering aims to partition the n observations

into k sets (k n) , S = {S1, S2, , Sk} so as to minimize
the within-cluster sum of squares. The default and efficient
distance measure in k-means is Euclidean distance, the
formula used in this algorithms is [9].
(4)
V.
METHODOLOGY
The process of this project system has illustrated in

Figure 4. We may summarize them in the following steps:
1) Image Test set selection and categorization.
An image test set has been selected subjectively, and
divided into six categories they are: cars, fruit, t-shirt,
motorcycles, and cats.
2) Image test set global features extraction.
The three global features has extracted from all image
test set and stored in images database.
3) Global features or texture features clustering by kmeans clustering algorithm.
4) Image test set specific features extraction.
The regions and its color has been determined and
bigger regions signature has been calculated and stored
each one with its color.
5)
Specific features clustering by k-means.
6) Target image selection.

7) Target image features clustering general and
specific each one alone.
8) Re-clustering the both types of target image
features with image test set information.
9) Matching and resulting
Matching and resulting will rely on the results of reclustering of target image features, this depend on a
weighting criteria witch give each feature the same weight.
Figure 7. Recall of global category.
The results has displayed as number of charts, since

experiments has carried on five image categories, these
image categories are varied of their retrieving results, the
categories was: cars, fruits, t-shirts, motorcycles and cats.
Figure 5. System GUI
VI.
RESULTS ANALYSIS
The graphical user interface is as depicted in Figure 5.

It has used exactly as the process above has listed.
A. Performance Measures
The performance measures used are: precision and
recall. The two most frequent and basic measures for
information retrieval effectiveness are precision and recall.
Precision
P=
# (relevant items retrieved)

_______________________
#(retrieved items)
Figure 8.
Global features category recall vs. precision.
1) Global features results

Global features performance measure, i.e. precision,
which has computed for each image category has illustrated
in Figure 6.
(5)
Recall
R =
# (relevant items retrieved)

_______________________
# (relevant items)
(6)
B. Results Analysis
Results of this research may divided into two parts:
Global Features results. And Local Features results.
Figure 9.
Precision of local category.
Recall is another performance measure of Global

Features, which, as well, computed per image category, the
results has illustrated in Figure 7.
Figure 6.
Precision of global category.
measure the similarity between images data set and target

image then displaying results with the measures of
performance, i.e. precision and recall.
Figure 10. Recall of local category..
The relationship accuracy graph between Precision and

Recall for Global Features has illustrated in Figure 8.
2) Local Features results
Local Features precision per image category has
illustrated in Figure 9.
Image retrieval using global features is giving good

results, but these results are not enough. Using color and
shape features as a local features help to improve the
retrieval mechanism, and it is better to exploit both local
and global features. Our experiments has given good
results due to using both of these two features, clustering
those using k-means algorithms. These results was different
based on image category that used. Some image categories
gives results better than others, this indicate the difficulty
and complexity of general CBIR, in the same time, it
encourage to experiment in specific domain CBIR, which
may give good ideas.
REFERENCES
[1]
[2]
[3]
[4]
[5]
Figure 11. Local category Recall vs. precision.
Local features recall per image category has illustrated

in Figure 10.
The relationship between Precision and Recall for local
features has illustrated in Figure 11.
3) Proposed system final results
Proposed system final result which consist of global
features and local features results as a graph has illustrated
in Figure 12.
[6]
[7]
[8]
[9]
Pavlidis, Theo. LIMITATIONS OF CONTENT-BASED IMAGE

RETRIEVAL. Invited Plenary Talk at the 19th Internat. Conf. on
Pattern Recognition, Tampa, Florida, e em e 2
on , u ui,
on ian
an , an
avi
a an en
S
- S
u time ia
n o mation
et ieva
an
ana ement Sp in e e in ei e e , 2 3
Kalantidis, Y., Spyrou, E., Mylonas, P., & Kollias, S
S
S
S
S
Semanti
e ia
aptation
an e sona ization S
, 2 1 t nte nationa
o ks op on
,2 1
Gangan, Manjary P., and R. Karthi. AUTOMATIC IMAGE
ANNOTATION BY CLASSIFICATION USING MPEG-7
FEATURES ,International Journal of Scientific and Research
Publications, Volume 2, Issue 9, September 2012.
Singh, N., Dubey, S. R., Dixit, P., & Gupta, J. P. SEMANTIC
IMAGE RETRIEVAL BY COMBINING COLOR, TEXTURE
AND SHAPE FEATURES. Computin S ien es
S , 2 12
nte nationa on e en e on
, 2 12
Su, a- wun ,
un- i uo, an
in ent S sen
- S
anu a omputin
, 2 12
nte nationa on e en e
on
, 2 12
Jasmine, K. Prasanthi, and P. Rajesh Kumar.
COLOR
HISTOGRAM AND DBC CO-OCCURRENCE MATRIX FOR
CONTENT BASED IMAGE RETRIEVAL, I.J. Information
Engineering and Electronic Business, 2014, 6, 47-54.
Gonzalez, Rafael C.DIGITAL IMAGE PROCESSING, Third
Edition, Pearson International Edition, 2003.
Han, Jiawei, Micheline Kamber, and Jian Pei. DATA MINING:
CONCEPTS AND TECHNIQUES. San Francisco, CA, itd:
Morgan Kaufmann 5 (2001).
Figure 12. Final results.
VII. CONCLUSION
CBIR was achieved using image semantic features,
which are image shapes, colors, and texture. We presents
an approach used clustering techniques in these features to

Ijcs 2016 0303012 PDF

Uploaded by

Copyright:

Available Formats

Ijcs 2016 0303012 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ijcs 2016 0303012 PDF

Uploaded by

Copyright:

Available Formats

International Journal of Computer Systems (ISSN: 2394-1065), Volume 03 Issue 03, March, 2016

An Image Retrieval Approach based on Colored Regions Features and a

Information Systems Department,

Content-based image retrieval (CBIR) is used computer

interprets easily, and the low-level features, which has used

Traditional Image retrieval could be done using

texture and color features. It has used images texture

The features that extracted are global features and then

Image standard deviation

Image retrieval approaches

Standard deviation is used usually to detect and distinct

In [3], Kalantidis et al, had presented a region-based

CBIR framework has consisted of three general steps

Image correlation with a referenced image.

Figure 2. Semantic gap in image retrieval

Shape invariant moment

Texture: std, entropy, correlation

Color & region features

Clustering region and color features

Clustering general features

Re-classify Target image texture

Re-classify Target image region and

Matching and resulting

Matching and resulting

Figure 4. Diagram illustrates the process of proposed system.

E. Shape color class

K-means algorithm for information clustering

k-means clustering aims to partition the n observations

The process of this project system has illustrated in

Specific features clustering by k-means.

6) Target image selection.

Figure 7. Recall of global category.

The results has displayed as number of charts, since

Figure 5. System GUI

The graphical user interface is as depicted in Figure 5.

# (relevant items retrieved)

Global features category recall vs. precision.

1) Global features results

# (relevant items retrieved)

Precision of local category.

Recall is another performance measure of Global

Precision of global category.

measure the similarity between images data set and target

Figure 10. Recall of local category..

The relationship accuracy graph between Precision and

Image retrieval using global features is giving good

Local features recall per image category has illustrated

Pavlidis, Theo. LIMITATIONS OF CONTENT-BASED IMAGE

Figure 12. Final results.

You might also like