Object Shape Representation by Kernel Density Feature Points Estimator

Natarajan Meghanathan, et al. (Eds): ITCS, SIP, JSE-2012, CS & IT 04, pp. 209–216, 2012.
© CS & IT-CSCP 2012 DOI : 10.5121/csit.2012.2119
Object Shape Representation by Kernel
Density Feature Points Estimator
Tranos Zuva1
, Oludayo O. Olugbara2
, Sunday O. Ojo3
and Seleman M.
Ngwira4
1, 4
Department of Computer Engineering, Tshwane University of Technology,
Pretoria, South Africa
zuvat@tut.ac.za
3
Faculty of Information and Communication Technology,
2
Department of Information Technology, Durban University of Technology,
Durban, South Africa
ABSTRACT
This paper introduces an object shape representation using Kernel Density Feature Points
Estimator (KDFPE). In this method we obtain the density of feature points within defined rings
around the centroid of the image. The Kernel Density Feature Points Estimator is then applied to
the vector of the image. KDFPE is invariant to translation, scale and rotation. This method of
image representation shows improved retrieval rate when compared to Density Histogram
Feature Points (DHFP) method. Analytic analysis is done to justify our method and we compared
our results with object shape representation by the Density Histogram of Feature Points (DHFP)
to prove its robustness.
KEYWORDS
Kernel Density Function, Similarity, Image Representation, Segmentation, Density Histogram
1. INTRODUCTION
With vast collection of digital images on personal computers, institutional computers and the
Internet, the need to find a particular image or a collection of images of interest has increased
tremendously. This has motivated the researchers to find efficient, effective and accurate
algorithms that are domain independent for representation, description and retrieval of images of
interest. There have been many algorithms that have been developed to represent, describe and
retrieve images using their visual features such as shape, colour and texture [1], [2], [3], [4], [5].
Visual feature representation and/or description play an important role in image classification,
recognition and retrieval. A successful image representation and description is dependent on the
selection of suitable image features to encode and quantification of these features [4].
Shape representation and description have been dominant in research area of image processing
because shape is considered to be the basis of human visual recognition [4]. The shape
representation can be classified as Region based or Contour based.

210 Computer Science & Information Technology (CS & IT)
The contour based techniques use the boundary of shape to describe an object. It is commonly
believed that human beings can differentiate objects by their boundaries or contours [2]. Usually
most objects form shapes with defined contours, making the use of these techniques most
appealing. The techniques can generally be applied to different application areas with a
considerable success. The techniques have a low computation complexity as compared to region
based techniques and they are sensitive to noise. These techniques in this group are well
described in [5].
The region based shape representation uses the boundary pixels and the interior pixels of the
shape (distribution of all pixels within contour). This group of shape representation algorithms are
robust to noise, shape distortion and they are applicable to generic shapes [6]. These techniques
can be found in [5].
This paper proposes a Kernel Density Feature Points Estimator representation of an image object.
This method imitates human visualization of image object shape and matching similar object
shapes. A comparison of retrieval of similar image object shapes is done between Kernel Density
Feature Points Estimator and Density Histogram of Feature Points representation of image object
shapes.
2. SHAPE REPRESENTATION BY KDFPE
This method describes the feature points within rings in an image grid. Assume we have a
silhouette object shape segmented by some means such as active contour without edges [7] and
let the feature points set ),( yxP (intensity function) of the object shape be defined as
Ν== εnniyxpyxP i ,.....,2,1),,(),( . (1)
The centroid of the object shape is calculated. The following formulae will be used to calculate
the centroid [8],[9]:
0,0
0,1
m
m
xc = (2)
0,0
1,0
m
m
yc = (3)
where 0,01,00,1 ,, mmm are derived from the silhouette moments given by
∑∑=
x y
ji
ji yxPyxm ),(, . (4)
The theorems that guarantee the uniqueness and existence of silhouette moments can be found in
[8]. For silhouette image ),( yxP , 0,0m the moment of zero order represents the geometrical area
of the image region and 1,00,1 ,mm moment of first order represents the intensity moment about
the y-axis and x-axis of the image respectively. The centroid ),( cc yx gives the geometrical
centre of the image region.

Computer Science & Information Technology (CS & IT) 211
Suppose the size of the grid occupied by the object shape is NxN. The vector dimension to
represent the density of object shape will be N-1. From the centroid we count the number of
image pixels in the rings with defined equal width around the centroid. The number of image
pixels in each ring is given as )......,,( 21 mi nnnX = where m is the number of rings from the
centroid.
The Kernel Density Feature Points Estimator (KDFPE) is then applied. The Second-order
Gaussian Kernel Density Estimator (SGKDE) is used.
The SGKDE is given in [10] as
2
1
2
1
2
11
)( ∑=





 −−
=
m
i
h
Xx i
e
mh
xf
π
(5)
The optimal bandwidth oh
for each image shape is calculated using the following second order
Gaussian plug-In formula [11]:
smho
5
1
059.1
−
= (6)
where m is the number of rings from the centroid and s is sample standard deviation.
The vector elements of the image are recalculated using the following:
2
2
1
2
1
)(







−
= o
i
h
X
o
i e
h
xf
π (7)
)( ixf becomes the image representation vector.
2.1. Illustration
Suppose object shape features are given on a grid as shown in figure 1.
0,0 1,0 2,0 3,0 4,0
0,1 1,1 2,1 3,1 4,1
0,2 1,2 2,2 3,2 4,2
0,3 1,3 2,3 3,3 4,3
0,4 1,4 2,4 3,4 4,4
Figure 1. Segmented object shape
The red-bold indicate the “on” pixels (pixels that belong to the image). The size of the grid
occupied by the object shape is 5X5. The centroid calculated by the two formulas above (2) and
(3) is (3, 2), the centroid pixel is in blue. The first rectangle boundary in figure 1 is made up of
the following pixels
(2,2)(2,3),(3,3),(4,3),(4,2),(4,1),(3,1),(2,1),

and there are seven “on” pixels that constitute our first element of the vector. The preliminary
vector representation of object shape in figure 1 is
1)8,(7,
The Kernel Density Estimator (KDE) is then applied. As mentioned earlier on the second-order
Gaussian KDE is going to be used. The sample points (in our example) with the unknown
distribution function are:
)1,8,7(
The Second-order Gaussian KDE is given as:
2
3
1
2
1
2
1
3
1
)( ∑=





 −−
=
i
h
Xx i
e
h
xf
π (8)
The optimal bandwidth oh for each image shape is calculated. Then we the recalculate the vector
elements of the image, using the following
2
2
1
2
1
)(







−
= o
i
h
X
o
i e
h
xf
π (9)
This is how the images will be represented.
3. SIMILARITY MEASUREMENT
To measure the similarity of the images we used the Cosine Coefficient given in [12] as
∑∑
∑
==
=
=
m
i
i
m
i
i
m
i
ii
QP
QP
s
1
2
1
2
1
cos (10)
The equation above is the normalized inner product and called the cosine coefficient because it
measures the angle between two vectors. It is often called angular metric.
4. ACCURACY MEASUREMENT
The accuracy of the system is measured by calculating the recall, the precision and effectiveness.
The following formulas were used [1]
databaseinreleventtotal
resultsquerytheinreleventtotal
N
A
recall ==
(11)
resultsquerytheintotal
resultsquerytheinreleventtotal
CA
A
precision =
+
=
(12)






≤
〉
=
NTif
T
A
NTif
N
A
esseffectiven
(13)
Where A is the number of relevant image objects retrieved, B is the number of relevant image
objects not retrieved, C is the number of irrelevant image objects retrieved andT is the user
required number of relevant image retrieval.
5. EXPERIMENTATION
The main objective is to find the effectiveness of KDFPE method and also compare it with other
representation methods. In this case a comparison is made with Density Histogram Feature
Points. The Cosine Coefficient Similarity measure is used in retrieval of similar image objects.
An image database of 200 shop items shapes is created. Some of the image objects are rotated at
90, 180 and 270 degrees. The images that are rotated were not rotated lossless, meaning
degradation of the image object occurred during rotation. The image objects were of different
dimensions MXN or NXN where M and N belong to real numbers when they are brought to the
system. The images that are used only have one image object with a homogeneous background.
The image object shape of a 45X45 grid is segmented. All images are converted to gray scale
images. They are then represented using KDFPE and DHFP. The average precision is calculated
for retrieval of images in the database using queries of images from the database. The average
effectiveness is calculated for retrieval images in the database using queries of images captured
using camera enabled devices. Matlab 7.6 was used to implement the system. Example of classes
of shapes experimented with are given in figure 2.
Figure 2. Examples of classes of items in database
In each class there are ten elements with some items rotated and scaled.

6. RESULTS
6.1. Similarity Matching
Different types of televisions are in the collection of televisions in the database. Some of the
segmented televisions are shown in figure 3 below. The KDFPE is capable of matching most
of televisions despite errors in segmentation, noise and distortion of the television shapes
due to transformations (scaling and rotation).
Figure 3. Segmented shapes that were considered similar by KDFPE
6.2. Comparison of Representation Methods (KDFPE and DHFP)
Figure 4. Comparison of KDFPE and DHFP
The results in figure 4 show that KDFPE is better in retrieving images that are in the database
using query images that belong to the database. It can be appreciated that the difference is not
very substantial when it comes to querying the database with images in it.
0
20
40
60
80
100
120
25 50 75 100
Precision%
Recall %
Comparison KDFPE & DHFP
Average Precision
KDHFP %
Average Precision
DHFP %

Figure 5. Comparison of effectiveness of the two methods
The results shown on figure 5 shows that KDFPE is better than the DHFP method when
using a query that does not belong to the database. The figure 5 shows a very substantial
difference in the effectiveness of the methods.
7. SUMMARY AND CONCLUSION
From the results it can be concluded that KDFPE method of image object representation is able to
differentiate similar object shapes just as human beings perceive image objects shapes. The
ability to calculate the optimal width seems to give KDFPE an advantage over DHFP method.
The KDFPE method shows that it is capable of retrieving images similar to images captured by a
camera enabled device as shown in figure 5. This characteristic is good in recommender systems
because people are looking for images similar to their own captured images but not necessarily in
the database. In future we are going to use different methods of calculating optimal width to
investigate if we may improve the effectiveness of KDFPE. Different kernel functions are going
to be used in future to investigate if they have any effect in the retrieval rate. It is important to
have a retrieval system that is capable of matching images that are in database with images that
are not in the database. KDFPE is capable of overcoming errors in segmentation and is robust to
segmentation noise.
REFERENCES
[1] M. X. Ribeiro, et al.,(2006) "Statistical Association Rules and Relevance Feedback: Power Allies to
Improve the Retrieval of Medical Images," Proceedings of the 19th IEEE Symposium on Computer-
Based Medical Systems.
[2] D. Zhang and G. Lu, (2004) "Review of shape representation and description techniques," Pattern
Recognition Society, vol. 37, pp. 1-19.
[3] Y. Li and L. Guan, (2006) "An effective shape descriptor for the retrieval of natural image
collections," in. Proceedings of the IEEE CCECE/CCGEI, Ottawa, pp. 1960-1963.
[4] X. Zheng, et al., (2007) "Perceptual shape-based natural image representation and retrieval," in
Proceedings of the IEEE International Conference on Semantic Computing, pp. 622-629.
0
20
40
60
80
100
120
25 50 75 100
Effectivenes%
Recall %
Comaprison KDFPE & DHFP
KDHFP %
DHFP %

[5] Y. Mingqiang, et al., (2008) "A survey of shape feature extraction techniques," in Pattern
Recognition, pp. 43-90.
[6] E. M. Celebi and A. Y. Aslandogan, (2005) "A comparative Study of Three Moment-Based Shape
Descriptors," Proceedings of the International Conference on Information Technology: Coding and
Computing.
[7] T. F. Chan and L. A. Vese, (2001) "Active Contours Without Edges," IEEE, vol. 10, pp. 266-277.
[8] J. Flusser, et al., (2009) "Moments and moment invariants in pattern recognition." West Sussex: John
Wiley & Sons Ltd.
[9] R. Mukundan and K. R. Ramakrishnan, (1998) "Moment functions in image analysis: theory and
applications." Singapore: World Scientic Publishing Co. Pte. Ltd.
[10] H. Shimazaki and S. Shinomoto, (2010) "Kernel bandwidth optimization in spike rate estimation," J
Comput Neurosci, vol. 29, pp. 171-182.
[11] J. S. Simonoff, (1996) "Smoothing Methods in Statistics," in Springer Series in Statistics, Springer,
Ed., ed.
[12] S.-H. Cha, (2007) "Comprehensive Survey on Distance/Similarity Measures between Probability
Density Functions," International Journal of Mathematical Models and Methods in Applied Sciences,
vol. 1, pp. 300-307.

Object Shape Representation by Kernel Density Feature Points Estimator

More Related Content

What's hot (17)

Similar to Object Shape Representation by Kernel Density Feature Points Estimator (20)

More from cscpconf (20)

Recently uploaded (20)

Object Shape Representation by Kernel Density Feature Points Estimator