IMAGE SEARCH USING SIMILARITY MEASURES BASED ON CIRCULAR SECTORS

Jan Zizka et al. (Eds) : ICAITA, SAI, CDKP, Signal, NCO - 2015
pp. 241–251, 2015. © CS & IT-CSCP 2015 DOI : 10.5121/csit.2015.51519
IMAGE SEARCH USING SIMILARITY
MEASURES BASED ON CIRCULAR
SECTORS
Jan Masek, Radim Burget, Lukas Povoda and Martin Harvanek
Department of Telecommunications, Faculty of Electrical Engineering,
Brno University of Technology, Brno, Czech Republic
masek.jan@phd.feec.vutbr.cz, burgetrm@feec.vutbr.cz,
xpovod00@stud.feec.vutbr.cz, xharva01@stud.feec.vutbr.cz
ABSTRACT
With growing number of stored image data, image search and image similarity problem become
more and more important. The answer can be solved by Content-Based Image Retrieval
systems. This paper deals with an image search using similarity measures based on circular
sectors method. The method is inspired by human eye functionality. The main contribution of the
paper is a modified method that increases accuracy for about 8% in comparison with original
approach. Here proposed method has used HSB colour model and median function for feature
extraction. The original approach uses RGB colour model with mean function. Implemented
method was validated on 10 image categories where overall average precision was 67%.
KEYWORDS
CBIR, circular sectors, cross-validation, image features, image processing, image similarity,
optimization
1. INTRODUCTION
Nowadays, the amount of transmitted image data through internet is every day still growing and
due to this fact digital image databases are filled with new terabytes of images. In order to search
and manage this data, there is strong need to index or categorize these images using proper
system. Searching images on the basis of similarity can be used in medicine, arts, industry [1],
security, military and many other areas [2].
This work deals with an image categorization and search on the basis of content. Systems that
provide this functionality are called Content-Based Image Retrieval (CBIR) [3]. These systems
search huge image databases, where for every image the special signature is created. The
signature is used for comparing with image we want to categorize. In our approach we improved
circular sector method introduced in [4] and we increased accuracy for about 8%.
CBIR systems usually use visual image properties like colour, texture and shape for creating
feature vectors that are saved in to the database. Visual image properties are compared by using
similarity measurements (Euclidean metrics, Manhattan metrics) and according to the value of

242 Computer Science & Information Technology (CS & IT)
measurements, images are compared or searched in database. CBIR systems use several methods
for the computing of feature vectors. Methods can be based on local or global feature extraction
or can be based on colour coherence vectors [5], colour moments [6], circular sectors [4] or
Gabor filters [6]. The CBIR system architecture is depicted in Figure. 1.
The main contribution of the paper is method that modifies original approach [4]. This approach
uses circular sectors method that is inspired by human eye functionality. We achieved higher
accuracy for about 8% when compared with [4]. We conducted parameter optimizations using
cross validation process and machine learning [19] to find optimal learning algorithm and its
configuration. Our approach uses different types of circular sector features where we used HSB
colour model with median function instead of RGB colour model with mean function for feature
computation.
The rest of this paper is organized as follows: The second section describes related work with
focus on CBIR systems. Section 3 describes circular sector method. In section 4 method
modification is described. Image data sets are described in section 5. The section 6 describes
optimization of parameters. Results are discussed in section 7 and section 8 concludes this paper.
Figure 1. Content base image retrieval system architecture.
2. RELATED WORK
Until today many content base image retrieval systems have been created [3]. We present several
leading systems in this chapter. For example QBIC system from IBM has been used for many
further work dealing with CBIR. Another leading systems are visualSeek or Netra [4]. From
these systems many following system have been derived [7], [8] and [9].
There are many works dealing with different image features. Histogram intersection computation
has been used to compare images in [10]. Cumulative histograms were described in [11] and

Computer Science & Information Technology (CS & IT) 243
spatial matching with colour histograms were described in [12]. In [13] and [14] is proven that
colour features are very suitable for similarity measurements.
We also described method based on dominant colours in [15] for measuring image similarity and
in [16] system for automatic image labelling using similarity measures is described. In [17] video
scenes were segmented using similarity measures.
3. CIRCULAR SECTORS METHOD
This method has been described in [4] and it is based on human eye principle. The human eye
firstly focuses on the center of image and then goes to the edges of image. The method creates
special image features that are obtained from image. Firstly, the center of image is determined
and then image is divided in to concentric circles Ci, where ݅ is number of circle. Then every
circle is divided to sectors Si, where Si = 8 Ci. In this case, seven circles are chosen and 252
sectors in whole image are created (see Figure 2 - left). Due to the fact that we use RGB color
model with 3 channels, we need to create 3·252 = 756 sectors. The next step is to compute mean
value in every sector (see Figure 2 - right). So the image feature is computed as mean value of
defined sector. When this method is applied on input image, output feature vector containing
mean values of all sectors is computed.
Figure 2. Circular sectors in the image (left), average colour values in each sector (right) [4].
To make this method rotation-invariant, the mean values of sectors are sorted in every circle.
Figure 3 shows that sorted sectors are similar when using normal or rotated image.
4. METHOD MODIFICATION
In originally described method authors used for feature extraction RGB channels and mean value
computation. We decided to create new image features using median function and using HSB
colour model. The comparison between these new features and previously used features will be

described in chapter 6. Our implementation of algorithm has been created in JAVA programming
language according to previous work [4] with our new modifications.
4.1 Median
The mean colour value of sector cannot exactly determine the distribution of pixel values (e.g. if
image contains a little noise). We modify previous method with using median values instead of
mean values.
4.2 HSB Colour Model
For human perception the HSB model suits better than RGB model. HSB is an abbreviation of
Hue, Saturation, and Brightness. This model use the cylindrical-coordinate representations of
values in an RGB model. We use the HSB model instead of RGB colour model.
Figure 3. Original image with original and sorted sector values (left), 30° rotated image with original and
sorted sector values [4].

5. DATA SETS
In this work we used same image data set as authors that described original circular sector
method [4]. This data set is available to download from [20]. Data set consists of 10 categories
(ancient, beach, bus, dinosaur, elephant, flower, food, horse, mountain, natives) where every
category contains 100 images. We have 1000 images overall. The images have dimension
354x256 pixels. The example of used images is shown in Figure 4.
Figure 4. Example of used images
6. PARAMETERS OPTIMIZATION
There are many options how to extract features from image. For example dimensions of image,
the number of circles for creating sectors. Features can also be extracted using RGB or HSB
colour model or computing median or mean value. We chose nine variants that we wanted to
compare. For every variant, features were generated to format suitable for RapidMiner [21] data
mining tool. This tool contains many machine learning algorithms (e.g. algorithms of artificial
intelligence, optimization algorithms). We used cross-validation process [19] (see Figure 5) that
computes accuracy for every variant. The cross validation process used SVM (Support Vector
Machines) algorithm [18] of artificial intelligence. The SVM algorithm had these parameters:
• SVM type: C-SVC
• Kernel type: linear
• C: 1.1
• Epsilon: 0.001
The results of cross-validation process for every variant is shown in Table 1. It shows that HSB
colour model has higher accuracy than RGB model and also median function achieves higher
accuracy than mean function. The best achieved accuracy is 75.6% for image with 400x400
pixels dimensions, with 7 circles and HSB model where features are computed using median
function. It also shows that our approach that uses HSB model with median function has higher

accuracy (75.6%) in comparison with original approach [4] that uses RGB model with mean
function. Our modified method achieves for about 8% higher accuracy.
Figure 1. The scheme of cross validation process in RapidMiner tool.
Table 1. Selected variants and their accuracy of classification.
Dimensions Circles RGB mean RGB median HSB mean HSB median
200x200 3 64.3 % 66.0 % 71.0 % 71.5 %
200x200 5 65.6 % 67.3 % 72.3 % 72.9 %
200x200 7 68.1 % 70.5 % 72.6 % 74.8 %
300x300 3 65.0 % 65.6 % 70.6 % 72.4 %
300x300 5 68.6 % 68.9 % 72.8 % 72.0 %
300x300 7 67.2 % 71.3 % 72.9 % 74.1 %
400x400 3 64.4 % 65.5 % 71.1 % 72.2 %
400x400 5 68.6 % 69.0 % 72.7 % 72.4 %
400x400 7 67.6 % 70.8 % 73.2 % 75.6 %
Table 2 shows confusion matrix for every image category. The best precision was achieved with
dinosaur category (97.09%) and the lowest precision was achieved category ancient (53.45%).
Table 2. Confusion matrix for parameters (dimensions 400x400, circles 7, HSB median).
Label (real values) Prec.
[%]ancient beach bus dinosaur elephant flower food horse mountain natives
Prediction
ancient 62 16 0 0 7 0 3 2 12 14 53.45
beach 11 62 3 0 1 0 5 2 20 2 58.49
bus 2 2 83 0 0 3 3 0 5 1 83.84
dinosaur 0 1 0 100 0 0 1 0 0 1 97.09
elephant 8 2 0 0 79 0 1 0 3 6 79.80
flower 0 0 4 0 0 88 4 0 0 2 89.80
food 0 3 4 0 0 7 72 1 2 11 72.00
horse 3 1 0 0 1 0 2 93 0 0 93.00
mountain 6 12 4 0 6 2 1 0 57 3 62.64
natives 8 1 2 0 6 0 8 2 1 60 68.18

7. RESULTS
We performed several comparison tests to verify our modified method. For evaluation, we used
precision that is computed:
ܲ =
ܰ୘୔
ܰ୘୔ + ܰ୊୔
where ܰ୘୔ is a number of true positive (relevant) images and ܰ୊୔ is number of false positive
(irrelevant) images. Firstly, one pattern image is selected and its feature vector is computed, then
this feature vector is compared with the feature vectors of all images from data set. When data set
contains 1000 images, the comparison process had to be executed 1000000 times.
Comparison has been done with computing Euclidean and Manhattan metrics
݀ாሺ‫,ݔ‬ ‫ݕ‬ሻ = ඩ෍ሺ‫ݔ‬௜ − ‫ݕ‬௜ሻଶ
ௗ
௜ୀଵ
݀ெሺ‫,ݔ‬ ‫ݕ‬ሻ = ෍|‫ݔ‬௜ − ‫ݕ‬௜|
ௗ
௜ୀଵ
where ݀ is the length of input feature vector and ‫ݔ‬ and ‫ݕ‬ are feature vectors of 2 images that are
being compared. For every image, ܰ the most similar images are selected, where we set ܰ =
ሼ10, 25, 50, 100ሽ and the precision is computed for ܰ images. Finally, the overall precision is
computed as average of all precisions computed for every image.
Table 3 shows precision of every category (each contains 100 images) using Euclidean metrics
and Table 4 shows precision using Manhattan metrics. Overall average precision is shown in
Table 5. The best achived precision was 67.23% for ܰ = 10 with using Manhattan metrics.
Table 3. Precision of every category using Euclidean metrics.
ܰ
Ancient
[%]
Mountain
[%]
Bus
[%]
Dinosaur
[%]
Elephant
[%]
Food
[%]
Horse
[%]
Beach
[%]
Flowers
[%]
Natives
[%]
10 44.3 51.7 62.5 97.7 65.2 56.9 89.8 53.9 72.2 43.3
25 33.48 44.44 52 97.28 52.48 44.68 82.24 44.04 59.16 33.92
50 30.004 40.58 42.4 94.8 44.72 36.54 73.04 37.56 43.96 28.52
100 25.73 34.26 34 80.44 37.76 28.79 56.24 31.51 29.95 24.81

Table 4. Precision of every category using Manhattan metrics.
ܰ
Ancient
[%]
Mountain
[%]
Bus
[%]
Dinosaur
[%]
Elephant
[%]
Food
[%]
Horse
[%]
Beach
[%]
Flowers
[%]
Natives
[%]
10 51 52 63.9 99.7 66 66 92.1 50.3 79.2 52.1
25 41.76 44.72 54.24 99.52 53.64 58 86.96 43.04 67.8 43.88
50 33.76 39.66 45.22 98.84 44.74 48.54 79.18 37.04 51.82 37.8
100 28.59 34.04 36.63 91.81 36.98 37.01 62.42 32.28 36.55 32.46
Table 5. Overall average precision
ܰ Euclidean distance Manhattan distance
10 63.75 % 67.23 %
25 54.57 % 59.36 %
50 47.22 % 51.66 %
100 38.25 % 42.74 %
All computations were performed on computer with processor Intel Core i5 2.5 GHz and with
4GB of RAM memory. The computing of feature vector for all images took 1 minute and 9
seconds. To find and compare input pattern image with all image feature vectors (1000) took
approximately 2 seconds.
The results of searching pattern image (see Figure 6) for horse category are shown in Figure 7.
When pattern image is rotated to left by 90°, the results (see Figure 8) contain 4 incorrectly
selected images.
Figure 6. Pattern image for horse category.

Figure 7. First 10 the most similar images of horse pattern image.
Figure 8. First 10 the most similar images of horse pattern image rotated about 90°
8. CONCLUSION
The main contribution of this paper is a method that increases accuracy in CBIR systems for
about 8% in comparison with original approach [4]. The origin achieved accuracy was 67.6%.
We are currently able to achieve 75.6% accuracy with using the same image data set. We tried to
find suitable parameters for circular sectors method. We selected the method because it is
inspired by human eye functionality. We conducted parameters optimization using cross
validation process with algorithms of artificial intelligence, where we found that HSB colour
model and median function for feature computation achieve better result than original approach
using RGB colour model with mean function for feature computation. For testing we used 1000
images from 10 categories. The best result of average precision was 67.23% with using
Manhattan metrics. The average time for image comparison with database was 2 seconds on
common computer.

ACKNOWLEDGEMENTS
Research described in this paper was financed by the National Sustainability Program under grant
LO1401 and by the Czech Science Foundation under grant no. 102/12/1274 and with MPO FR-
TI4/151, Czech Republic. For the research, infrastructure of the SIX Center was used.
REFERENCES
[1] A. Rangkuti, Haris, et al. Analysis of Image Similarity with CBIR Concept Using Wavelet Transform
and Threshold Algorithm. In: Computers & Informatics (ISCI), 2013 IEEE Symposium on. IEEE,
2013. pp. 122-127.
[2] E. Chalom, Asa Eran, Biton, Elior. Measuring image similarity: an overview of some useful
applications. IEEE Instrumentation & Measurement Magazine, IEEE, 2013, vol. 16, no. 1, pp. 24-28.
[3] F. Long, H. Zhang and D. Dagan Feng., "Fundamentals of Content-Based Image Retrieval".
Multimedia Information Retrieval and Management–Technological Fundamentals and Applications,
Springer-Verlag, pp. 1-26, 2003.
[4] Omar, Samia G., Ismail, Mohamed A.; Ghanem, Sahar M. WAY-LOOK4: A CBIR system based on
class signature of the images’ color and texture features. In: Computer Systems and Applications,
2009. AICCSA 2009. IEEE/ACS, International Conference on. IEEE, 2009. pp. 464-471
[5] Pass, Greg, Zabih, Ramin, Miller, Justin. Comparing images using color coherence vectors. In:
Proceedings of the fourth ACM international konference on Multimedia. ACM, 1997. pp. 65-73.
[6] Singh, S. Mangijao, Hemachandran, K Content-Based Image Retrieval using Color Moment and
Gabor Texture Feature International Journal of Computer Science Issues (IJCSI), 2012, vol.9, no.5, s.
299-309.
[7] Y. Rui and T. S. Huang, "Image Retrieval: Current Techniques, Promising Directions, and Open
Issues", Journal of Visual Communication and Image Representation 10, pp. 39–62, 1999.
[8] T. Wang, Y. Rui, J. Guang, Sun, "Constraint Based Region Matching for Image Retrieval",
International Journal of computer vision 56 1/2, pp. 37-45, 2004.
[9] A. M. W. Smeulders, M. Worring, S. Santini, A. Gupta, and R. Jain, "Content-based image retrieval
at the end of the early years", IEEE Transactions On Pattern Analysis and Machine Intelligence, Vol.
22, No. 12, pp. 1349-1380, 2000.
[10] Michael J. Swain and Dana H. Ballard, “Color indexing,” International Journal of Computer Vision,
vol. 7, no. 1, pp. 11–32, June 1991.
[11] Markus Stricker and Markus Orengo, “Similarity of color images,” in In Proceedings of SPIE Storage
and Retrieval for Image and Video Databases, 1995, pp. 381–392.
[12] Markus Stricker, Alexander Dimai, and Er Dimai, “Color indexing with weak spatial constraints,” in
In Proceedings of SPIE Storage and Retrieval for Image and Video Databases, 1996, pp. 29–40.
[13] Gal Chechik, Varun Sharma, Uri Shalit, and Samy Bengio, “An online algorithm for large scale
image similarity learning,” in Advances in Neural Information Processing Systems, 2009.
[14] Gal Chechik, Varun Sharma, Uri Shalit, and Samy Bengio, “Large scale online learning of image
similarity through ranking,” Journal of Machine Learning Research, vol. 11, pp. 1109–1035, March
2010.
[15] J. Karasek, R. Burget, V. Uher, J. Masek, M. Dutta, Color Image (Dis) Similarity Assessment and
Grouping based on Dominant Colors. In 2014 37th International Conference on Telecommunications
and Signal Processing (TSP).Berlin, Germany: 2014. pp. 631-634. ISBN: 978-80-214-4983- 1.
[16] V. Uher, R. Burget, J. Karasek, J. Masek, M. Dutta, M. Automatic Image Labelling using Similarity
Measures. InMEDCOM 2014 CD-ROM. Greater Noida: IEEE, 2014. pp. 101-104. ISBN: 978-1-
4799-5096- 6.
[17] R. Burget, K. Ray, V. Uher, J. Masek, M. Dutta, Supervised Video Scene Segmentation using
Similarity Measures Supervised Video Scene Segmentation using Similarity Measures. In 36th
International Conference on Telecommunications and Signal processing. 2013. pp. 793-797. ISBN:
978-1-4799-0402- 0.

Computer Science & Information Technology (CS & IT)
[18] Chang, Chih-Chung, Lin, Chih
Transactions on Intelligent Systems and Technology (TIST), 2011, vol. 2, no. 3, s. 27.
[19] Akthar, Fareed, Hahne, Caroline. RapidMiner 5: Operator Reference. Dortmund: Rapid
2012. 990 s
[20] http://wang.ist.psu.edu/docs/related.shtml
[21] https://rapidminer.com/
AUTHORS
Jan Masek is Ph.D. student at the Department of
Electrical Engineering, Brno University of Technology, Brno, Czech Republic. He
obtained his MSc. in 2012 (Communications and Informatics). He is interested in image
processing, data mining, parallel systems.
Dr. Radim Burget is associated professor at the Department of Telecommunications,
Faculty of Electrical Engineering, Brno University of Technology, Brno, Czech Republic.
He obtained his MSc. in 2006 (Information Systems) and his finished his Ph.D. in 2010.
He is associated professor since 2014. He is interested in image processing, data mining,
genetic programming and optimization.
Lukas Povoda is Ph.D. student at the Department of Telecommunications, Faculty of Electrical
Engineering, Brno University of Technology, Brn
(Communications and Informatics). He is interested in image processing, text processing, and genetic
programming.
Martin Harvanek obtained his MSc. in 2014 at the Department of Telecommunications, Faculty o
Electrical Engineering, Brno University of Technology, Brno, Czech Republic. He is interested in image
processing and data mining.
Computer Science & Information Technology (CS & IT)
Chung, Lin, Chih-Jen. LIBSVM: a library for support vector machines. ACM
Transactions on Intelligent Systems and Technology (TIST), 2011, vol. 2, no. 3, s. 27.
Akthar, Fareed, Hahne, Caroline. RapidMiner 5: Operator Reference. Dortmund: Rapid
http://wang.ist.psu.edu/docs/related.shtml
Jan Masek is Ph.D. student at the Department of Telecommunications, Faculty of
Electrical Engineering, Brno University of Technology, Brno, Czech Republic. He
obtained his MSc. in 2012 (Communications and Informatics). He is interested in image
processing, data mining, parallel systems.
et is associated professor at the Department of Telecommunications,
Faculty of Electrical Engineering, Brno University of Technology, Brno, Czech Republic.
He obtained his MSc. in 2006 (Information Systems) and his finished his Ph.D. in 2010.
ted professor since 2014. He is interested in image processing, data mining,
genetic programming and optimization.
Engineering, Brno University of Technology, Brno, Czech Republic. He obtained his MSc. in 2014
Martin Harvanek obtained his MSc. in 2014 at the Department of Telecommunications, Faculty o
251
Jen. LIBSVM: a library for support vector machines. ACM
Akthar, Fareed, Hahne, Caroline. RapidMiner 5: Operator Reference. Dortmund: Rapid-I GmbH,
o, Czech Republic. He obtained his MSc. in 2014
Martin Harvanek obtained his MSc. in 2014 at the Department of Telecommunications, Faculty of

IMAGE SEARCH USING SIMILARITY MEASURES BASED ON CIRCULAR SECTORS

More Related Content

What's hot (19)

Similar to IMAGE SEARCH USING SIMILARITY MEASURES BASED ON CIRCULAR SECTORS (20)

More from cscpconf (20)

Recently uploaded (20)

IMAGE SEARCH USING SIMILARITY MEASURES BASED ON CIRCULAR SECTORS