Improved CBIR System using Edge Histogram

Descriptor (EHD) and Support Vector Machine

(SVM)
Pradnya Vikhar Pravin Karde
Research Scholar, Head, Department of Computer Engineering,
Shri Sant Gadge Baba Amravati University, Government Polytechnic, Amravati
Amravati, India India
[email protected] [email protected]

Abstract- Due to widespread applications found in many observed that it is a very expansive and subjective task. This
areas, Content Based Image Retrieval (CBIR) system is limitation explores the new area of Content Based Image
attracting attention of many researchers. Effectiveness of any Retrieval (CBIR) system for searching and retrieval of images
CBIR system depends on the features extracted to represent an [1][3]. In CBIR system visual contents, such as color, texture
image. So feature extraction is the crucial step in design and
development of any Content Based Image Retrieval system. Most
and shape takes the place of textual annotation to represent an
commonly used features to represent images are Color, texture image. Since 1990s, it is persistent area of research and
and shape. Recently developed CBIR system combines these development. Many researchers carried out extensive efforts to
features to effectively represent an image. develop and design efficient and effective CBIR systems. The
This paper first discusses the concept and scope of content based major work done in this area is focused on feature extraction
image retrieval system. It also includes the overview of MPEG-7 techniques to represent the images, similarity matching to
edge histogram descriptor (EDH) to extract the contents from compare the feature vectors obtained from feature extraction,
images. Further it gives the idea of Support Vector Machine indexing to list the output images based on certain attributes,
(SVM) classifier. In this paper the basic CBIR system is and relevance feedback which take users view to display the
developed by combining features like color moments, color-
correlogram and Gabor texture features along with edge
output[1][2].
histogram descriptor. Further the results obtained are compared The main focus of the paper is feature extraction using MPEG-
with CBIR system using SVM classifier. 7; Edge Histogram Descriptor (EHD) [4]. It is a standard
descriptor used to represent texture features of an image. To
Keywords: CBIR, Feature extraction, MPEG-7, edge histogram
descriptor (EDH), Support Vector Machine (SVM)
effectively represent image contents, this technique is further
combined with color moment, color auto-correlogram and
I. INTRODUCTION Gabor texture features. Further to improve overall performance
Because of rapid development in digital technology, enormous of CBIR system, the SVM classifier is used. [5][6]. It can work
amount of digital information in the form of images and videos well when the dimension of the input space is very high and
is generated every day. Also rapid transmission of the digital even without any prior knowledge. So it is considered as a
information is required for various purposes [1][2]. Now a good classifier.
day’s technology demands the use of digital information in II. CONTENT BASED IMAGE RETRIEVAL SYSTEM
form of images in various areas of life including security,
medicine, bioinformatics, amusement, designing, Content based image retrieval (CBIR) systems involves
advertisements and engineering sciences [1][2]. There is a need applications which includes automatic indexing, searching,
of proper mechanisms and tools for effective use of such vast retrieving and browsing of images from image database. It
digital information. It leads to the development of Content uses own visual contents of an image such color, shape and
Based Image Retrieval System (CBIR). texture to search large image database as per users’ interest
In early days, textual annotation was used to represent image. [1][2]. A typical steps involved in design of any CBIR system
Then the images were searched and retrieved using text-based is as shown in figure 1 [2]. In CBIR system, user gives the
concept from traditional database management system. As the input query in the form of image. First feature extraction of
approach is based on text-based search, it is commonly input query image and images present in database is done using
acknowledged as Text-based Image Retrieval (TBIR) [3]. The some feature extraction techniques. Then it generates the
problems in front of Text-Based image retrieval (TBIR) system feature vectors of query images and the images present in
became more and more serious with rapid generation of digital image database. It forms the features space. Similarity between
images in various fields of entertainment, science and feature vector of query image and feature vectors of images
engineering. The main difficulty faced by this approach is to present in feature database is calculated. After similarity
describe rich contents in images manually. Obviously, it is matching, most similar images are displayed as an output. In
978-1-5090-5515-9/16/$31.00 ©2016 IEEE
CBIR system output is not a single image but it is the list of 4) By applying above steps, from every sub-image 5-bin edge
similar images [7]. histogram is obtained. Five bins are vertical, horizontal, 450
diagonal, 1350 diagonal, and no-direction
5) Then, the value of each bin in the sub-image is computed by
User Query Image Database normalizing total number of image blocks in the sub-image.
6) Finally for these normalized bins nonlinear quantization
value is calculated to limit the number of bits sufficient for
Feature Feature the descriptor.
Extraction Extraction Texture features of input image and images present in database
are extracted using above steps.

Feature Feature Database

Vector

Feature Space

1 -1 1 1 √2 0 0 √2 2 -2
1 -1 -1 -1 0 - √2
-√2 0 -2 2
Similarity
Comparison
Fig 2: Five types of edges and corresponding filters [9]

IV. SUPPORT VECTOR MACHINE (SVM)

Output The basic idea of SVM is shown in Figure 3[5][6].The data
Images points are identified as being positive or negative. The hyper-
plane which separates the data points by a maximal margin is
identified. Following Figure only shows the two-dimensional
Fig1: A typical view of CBIR System
case where the data points are linearly separable [5][6].

III. EDGE ORIENTATION HISTOGRAM Positive

Instances
Edge Histogram Descriptor (EHD) belongs to the family
MPEG-7 descriptors. MPEG-7 provides different standard
Multimedia Descriptors to describe and interpret visual
contents for image, audio and video retrieval [8]. It offers the
interoperability across different multimedia databases. Main
Negative
visual features descriptors are color, shape and texture Instances
descriptors. In proposed methodology, Edge Histogram
Descriptor (EHD) is applied to extract texture features of
images. Maximizes distance to
The EHD [4][8] characterizes edges to represent spatial nearest points
distribution in an image. Feature extraction process using EHD
consists of following steps [4][8]: Fig 3: Data classification using support vector machine [6]
1) First, the array of digital image is spitted into equal 4X4
subparts/ sub-images. The mathematical formulation of problem to be solved using
2) In next step, every subpart is further divided into non- SVM is as following [5][6]-
1
overlapping square blocks. Here size of blocks are depends MIN || ||
on the resolution of input image. , 2

3) From every block the edge is calculated and then type is

identified using the filter coefficients illustrated in figure 2. . . = +1 ⇒ . + ≥ +1
There are six types of edges. The type of edge may be
vertical, horizontal, 450 diagonal, 1350 diagonal, no- . . = −1 ⇒ . − −1
direction edge and no-edge.
Initially first five types of edges are identified and no-edge . . ( . + )≥ ∀
blocks can be automatically obtained after the process of
normalization. The identification of each data point xi is yi, which can take a
value of +1 or –1 (representing positive or negative
respectively). The equation to calculate hyper-plane is as Step 5: In this step Color Moment [2] up to 3rd order is
follows calculated. It represents distribution of color in images. A color
= . ̅+ distribution of image is efficiently and effectively represented
The scalar b is also termed as the bias. A standard method to by using color moments.
solve this problem is to apply the theory of Lagrange to
convert it to a dual Lagrangian problem. The dual problem is First order (mean) :
as the following = ,
1
min Ψ ( ) = min . −
2 Second order (variance):

= − ,

=0
Third order (skewness):
≥ 0, ∀ 1
Ѕ = −
The variables are the Lagrangian multipliers for
corresponding data point xi.
Step 6: In this step, the widely used Gabor wavelet [2] method
V. METHODOLOGY to extract image texture feature, is used to compute mean
In this paper, Edge Histogram Descriptor (EHD) of MPEG-7 is amplitude and mean squared energy. Many approaches are
used to extract texture features of images [9][10. To represent recommended to interpret texture of images by using Gabor
image more effectively along with texture, color features i.e. filters. In most CBIR systems are developed using Gabor
color auto- correlogram, color moments are also used. Thus wavelet, feature vector which is constructed by calculating
the approach combines color auto- correlogram, color moments mean and standard deviation of distribution of wavelet
and Gabor wavelet transforms along with EHD to develop transform coefficients. A two dimensional Gabor function
basic Content Based Image Retrieval system. The proposed g(x,y) to represent texture features of image is defined as:
methodology to develop a complete CBIR system contains
following steps: 1 1
( , )= exp − + +2
2 2
Step 1: Initially all the images are stored in a directory, called
image database. First step is to select directory of images.

Step 2: All images in database are resized using the process of Where and are the standard deviations along the x and y
normalization (set size of all images to 384 x 256). direction.
Step 7: By this step, all feature vectors calculated above are
Step 3: Apply above mentioned steps to compute Edge
combined to form the feature vector. Thus EHD, Color auto
Histogram Descriptor (EDH) to represent the texture features
Correlogram, Color Moments, Mean amplitude, Mean Squared
on images [4][8].
Energy, and wavelet moments of an image are combined.
Step 4: Then compute Color Auto Correlogram [2] as follows- Step 8: In this step feature database is generated. Steps 2 to 7
Consider an image (I), let the size n x m. and distance (d) is ϵ are applied to each image in database to calculate feature
{1, 2,.....,n}, then for image I; correlogram for i, j ϵ {1,2,.....,m} vector and feature vectors of all images in database form the
and k ϵ {1,2,.....d} is defined as feature database.

( ) Step 9: Through interface developed, user can select Query

,  ∈ , ∈ [ ∈ |[ − ]= ] image as an input.
Step 10: Using steps from 2 to 6, now one can generate feature
The color correlogram is characterized by a tabled index with vector for input query image (again using EHD, Color auto
pair of colors (ci,cj); where probability of finding a pixel with Correlogram, Color Moments, Mean amplitude, Mean Squared
color j at a distance k from a pixel with color i in the image, is Energy, and wavelet moments).
determined by the kth entry for (i, j). To develop image
retrieval systems using Color Correlogram has an advantage, as Step 11: Next step is to match feature vector of query image
it makes the system more robust against large alteration in and feature vectors stored in feature database (by step 8). Many
appearance of the same picture. measures are available to compute the similarity between the
images. Our approach uses Euclidean distance [2] method to
compute similarity distance. It is calculated as follows-
Consider two feature vectors a and b. suppose a = (a1, a2, .....,
an)T and b = (b1, b2,....., bn)T, the distance using Euclidean
method is calculated as follows:

= ( − )

Step 12: Similar images retrieved in step 11 are sorted based

on similarity distance and most similar images are displayed as
an output.
Step 13: After calculating feature vector using steps 2-6, apply
SVM classifier on the feature vectors. The classification of
feature vectors is done by SVM. Then similarity matching of
feature vectors as per step 11 is done, which will display the
resultant images.
Step 14: The comparison of systems before and after applying
SVM is performed using precision and recall measures
[1][2][7]. Fig 5. Query Image and Output generated after applying SVM Classifier

VI. RESULT AND DISCUSSION Recall and precision are widely used measures to calculate the
The approach discussed to develop CBIR system is performance of CBIR system. The results obtained before and
implemented in MATLAB. The interface is developed which is after applying the SVM technique are calculated and
used to load the feature database and to select the input image. represented using recall and precision graph. It is shown in
User can select query image whose feature vector calculated figure 6 and figure 7.
which is further compared with vectors in feature database. By In the graphs X- axis represents number of input queries and
computing similarity distance, the output is displayed. The Y-axis represents values of recall and precision. Recall is
system developed is tested on standard IMAGEVARY dataset shown by blue color line and precision is shown by red color
containing 1000 images. After giving query image as an input, line.
the system generates an output after indexing (sorting). It is By comparing graph from figure 6 (before applying SVM) and
seen that the system, after combining many features like EHD, figure 7 (after applying SVM), it is proved that using SVM the
color auto-correlogram, color moment and Gabor wavelet performance of Content Based Image Retrieval system is
transform generates the good output which is shown in figure improved significantly.
4.
recall
Further classification of feature vectors is done using SVM 0.45
classifier and results are displayed after similarity calculations. precision
0.40
It is shown in figure 5.
0.35
0.30
Values

0.25
0.20
0.15
0.10
0.05
0.00
1 2 3 4 5 6 7 8 9 10

Input Query

Fig 6. Recall and Precision graph before applying SVM

Fig 4.Query Image and Output generated before applying SVM Classifier
 [7] Vikhar Pradnya A.. "Content-Based ImageRetrieval (CBIR): State-of-the-
0.80 Recall Art and Future Scope for Research", IUP Journal of Information
Technology, Vol. 6,Issue 2, pp. 64-84, June 2010
0.70 Precision
[8] Laila Shoukry, Sebastian Klenk, Gunther Heidemann, “MPEG-7 Feature
Visualization for CBIR Systems”, in proceedings of International
0.60 Conference on Computer Theory and Applications (ICCTA 2010), 2010
[9] M H Saad, H I Saleh, H Konbor, M Ashour, “Image Retrieval based on
Values

0.50
Integration betweenYCbCr Color Histogram and Shape Feature”, IEEE
0.40 Computer Engineering Conference (ICENCO), 2011, vol., no., pp.97-
102, 27-28,2012
0.30 [10] Ching-Hung Sua, Mohd Helmy Abd Wahab and Tsai-Ming Hsieh,
“Image Retrieval based on color and texture features”, IEEE 9th
0.20 International Conference on Fuzzy Systems and Knowledge Discovery
(FSKD 2012), PP 1816-1819, 2012
0.10
0.00
1 2 3 4 5 6 7 8 9 10
Input Query

Fig 7. Recall and Precision graph after applying SVM

VII. CONCLUSION
This paper focuses on main issue of feature extraction in
development CBIR system. Feature extracted should be able to
effectively represent and interpret the rich contents of an image
in a database. If one combines number of features together to
generate the feature space, the image content can be described
and represented more effectively. Hence it will consequently
results in effective CBIR system. The proposed methodology
presented here uses MPEG-7 Edge Histogram Descriptor
(EHD) feature extraction technique and further it is combined
with Color auto-correlogram, color moment and Gabor wavelet
transform to effectively represent the image. To improve the
performance of CBIR system SVM classifier is suggested
which had considerably increased the performance of the
system.
REFERENCES
[1] Suman Lata, Parul Preet Singh, “A Review on Content Based Image
Retrieval System”, International Journal of Advanced Research in
Computer Science and Software Engineering, Volume 4, Issue 5, PP
561-566, May 2014
[2] Shereena V.B, Julie M. David, “Content Based Image Retrieval: A
Review”, Computer Science & Information Technology, Computer
Science Conference Proceedings (CSCP), pp. 65–77, 2014
[3] Shiv Raj Singh, Dr. Shruti Kohli”, Enhanced CBIR using Color Moments,
HSV Histogram, Color Auto Correlogram, and Gabor Texture”,
International Journal of Computer Systems (ISSN: 2394-1065), Volume
02– Issue 05, pp 161-165, May, 2015
[4] Jun Zhang, Lei Ye, and Jianhua Ma, “MPEG-7 Visual Descriptors and
Discriminant Analysis”, The Handbook of MPEG Applications:
Standards in Practice, John Wiley & Sons Ltd, 2011
[5]Sujata T Bhairnallykar, V.B.Gaikwad, “ Content based Medical Image
Retrieval with SVM Classification and Relevance Feedback”,
International Journal of Applied Information Systems (IJAIS) – ISSN :
2249-0868, PP 25-29, 2013
[6] V. Karpagam, R. Rangarajan, “Improved content-based classification and
retrieval of images using support vector machine”, Research
communications, current science, VOL. 105, NO. 9, PP 1267-1275,
November 2013.