Madireddy et al.

SpringerPlus 2014, 3:674

http://www.springerplus.com/content/3/1/674
a SpringerOpen Journal

RESEARCH Open Access

A modified shape context method for shape

based object retrieval
Radhika Mani Madireddy1, Pardha Saradhi Varma Gottumukkala2, Potukuchi Dakshina Murthy3 and
Satyanarayana Chittipothula4*

Abstract
The complexity in shape context method and its simplification is addressed. A novel, but simple approach to
design shape context method including Fourier Transform for the object recognition is described. Relevance of
shape context, an important descriptor for the recognition process is detailed. Inclusion of information regarding all
the contour points (with respect to a reference point) in computing the distribution is discussed. Role of similarity
checking the procedure details regarding the computation of matching errors through the alignment transform are
discussed. Present case of shape context (for each point with respect to the centroid) descriptor is testified for its
invariance to translation, rotation and scaling operations. Euclidean distance is used during the similarity matching.
Modified shape context based descriptor is experimented over three standard databases. The results evidence the
relative efficiency of the modified shape context based descriptor than that reported for other descriptor of
concurrent interests.
Keywords: Object recognition; Shape representation; Feature extraction; Distance measure

Introduction compounded image descriptor (Li and Lee 2005), shape

Although significant progress is witnessed in the field of matrix (Goshtasby 1985), the grid technique (Lu and
automated object recognition, it is still remains challen- Sajjanhar 1999) and shock graph (Sebastian et al. 2004;
ging task (Zhang and Lu 2004; Iyer et al. 2005) from the Siddiqi et al. 1999) etc. However, the contour based rep-
broad purview of machine learning and computer vision resentation is reported to be more efficient (Yang et al.
processes of contemporary requirements. The shape of 2008). Several recently reported contour based methods
an object contains (Forsyth and Mundy 1999) an im- rely on viz., Fourier transform (Zahn and Roskies 1972;
portant, unique and characteristic features of the object. Wallace and Wintz 1980; Kunttu et al. 2006), curvature
The shape based methods consider either the contour scale space (CSS) (Mokhtarian and Mackworth 1986;
or the entire region of the object. The consideration of Abbasi et al. 1999, 2000), wavelet transform (Chauang
contour involves less representative points in compari- and Kuo 1996; Yadav et al. 2007), contour displacement
son with the region based methods (Nixon and Aguado (Adamek and O’Connor 2004), chain codes (Junding
2002). The region-based methods consider the global and Xiaosheng 2006), autoregressive (Dubois and Glanz
information (all the pixels within a shape) for the design 1986), Delaunay triangulation (Tao and Wi 1999), multi-
of the descriptor which involves the geometrical mo- resolution polygonal (Day et al. 2004) robust symbolic
ments (Hu 1962; Flusser 2000), Zernike moments (ZM) representation (Daliri and Torre 2008), distance sets
(Teague 1980; Khotanzad 1990), pseudo- Zernike mo- (Grigorescu and Petkov 2003), elastic matching (Attalla
ments (Belkasim et al. 1991), Legendre moments (Teague and Siy 2005) etc techniques for the design of the shape
1980), and Tchebichef moments (Mukundan et al. 2001), descriptor. Basing on the consideration of shape bound-
generic Fourier descriptor (FD) (Zhang and Lu 2002), aries (Petrakis et al. 2002; Arica and Vural 2003; Bartolini
et al. 2005; Lateckia et al. 2005; Alajlan et al. 2007), dy-
* Correspondence: [email protected]
namic programming (DP) technique is also adopted to
4
Department of CSE, Jawaharlal Nehru Technological University Kakinada, achieve high accuracy rate. The DP based techniques suf-
Kakinada, AP 533003, India fer from being computationally expensive and get reduced
Full list of author information is available at the end of the article

© 2014 Madireddy et al.; licensee Springer. This is an Open Access article distributed under the terms of the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction
in any medium, provided the original work is properly credited.
Madireddy et al. SpringerPlus 2014, 3:674 Page 2 of 12
http://www.springerplus.com/content/3/1/674

to be impractical for large databases, despite the fact that The phase wise information during the processing is
they offer better performance. presented in section-Design of system, while the proposed
Generally, the descriptor relevant to the shape context indices to estimate the performance are presented in
(Belongie et al. 2002) method for object recognition is de- section-Performance.
veloped with an established correspondence between the
point sets. The procedure combines the shape context in- Design of system
formation with the information formatted by using thin The details of various stages involved with the proposed
plate spline (Bookstein 1989) processing. Due to the proven object recognition by using shape context are schematic-
simplicity and capability of discrimination, the shape con- ally depicted in Figure 1. The proposed system consists
text based methods proficiently proposed in the literature of four successive steps viz:
(Dubois and Glanz 1986; Tao and Wi 1999; Day et al. 2004;
Daliri and Torre 2008; Grigorescu and Petkov 2003; Attalla (i) Shape representation with contour
and Siy 2005; Petrakis et al. 2002; Arica and Vural 2003; (ii) Computation of Shape Context
Bartolini et al. 2005; Lateckia et al. 2005; Alajlan et al. (iii) Construction of histogram for each bin of shape
2007; Belongie et al. 2002; Bookstein 1989; Mori and Malik context
2003; Thayananthan et al. 2003; Zhang and Malik 2003; (iv) Shape description by using Fourier Transform
Salve and Jondhale 2010). Recently, Xin Shu proposed Con-
tour Points Distribution Histogram (CPDH) (Shu and Xiao The descriptor is further expected to a training stage
Jun 2011) for the shape context method. The shape match- viz., shape toning and ranking. An overview of all these
ing process which speaks out the performance of a descrip- stages of processing implies that the shape based object
tor is dealt in different ways. The Zucker et al (Siddiqi et al. recognition system includes the salient features of stages,
1999) has developed shock graph grammar and the relevant such as shape representation, shape description and shape
tree matching algorithm. The spectral distance (based on toning. Contour based shape representation is considered
diffusion geometry, heat trace) estimated through the as the initial step of processing. The second step includes
Laplacian transform is also used for matching (Bronstein description of the shape representation points. Belongie
and Kokkinos 2010; Bronstein and Bronstein 2011; Shape Context (BSC) (Belongie et al. 2002) is popularly
Konukoglu et al. 2013). On the other hand, the Fourier used method for describing the shape of the object.
transform based matching procedures are is also popular During the second step, the contour of the given object is
(Cem Direko glu and Nixon 2011; Xingyuan and Zongyu described (Figure 2(a)) by constructing the BSC. During
2013; Ghazal et al. 2009; Ghazal et al. 2012). the construction of BSC, the angle between any two points
In the wake of the results reported in the area of shape (on log polar transform) is measured with respect to con-
context based object recognition techniques involving a stant center on x axis given by the Equation-(1).
wide variety of design of description and matching mea-

y
sures, it serves that the utility of the Fourier based descrip- θðx; y Þ ¼ tan−1 2 ð1Þ
tors for the shape context based recognition presents a y1
superior method rather than the contour based methods.
where:
Hence, the authors propose for the design of a novel hybrid
contour based shape descriptor which is constructed with
θ(x, y) is the angle measurement between two points x
respect to the centroid, while the feature vector is esti-
and y,
mated by a 1D Fourier transform. The shape toning phase
y2 is the y coordinate of the first point,
is involves the Euclidean Distance to enhance the quality.
y1 is the y coordinate of the second point.
The paper is organized in three sections. Introduction to
the computerized object recognition method is presented
To test the invariance property of the BSC, a Modified
in section-Introduction. Methodology adopted for the
Shape Context (MSC) is presently proposed (Figure 2(b)).
present shape context technique is presented in section-
The MSC measures the angle between any two points
Methodology along with the information for indices to
with respect to the centroid. If the total no. of contour
evaluate its performance. The results obtained by adopting
points are Z then the farthest two points will be selected
present method to the standard databases and their trends
and the angle between these points is measured by the
are presented in section-Results and discussion along with
Equation-(2).
the relevant discussions of performance.


‐1 m2 ‐m1
Methodology θðx; y Þ ¼ tan ð2Þ
1 þ m 2 þ m1
A multi staged novel and hybrid shape context based
scheme for the object recognition process is proposed. where:
Madireddy et al. SpringerPlus 2014, 3:674 Page 3 of 12
http://www.springerplus.com/content/3/1/674

Figure 1 Schematic diagram of the proposed system with shape context.

θ(x,y) is the angle measurement of a point (x, y), By using EAL, the representation of the contour is re-
m1 is the slope of the line between first point and stricted to N-number of points. The proposed method
second point, uses EAL method to sample the finite number of con-
m2 is the slope of the line between first point and tour points. For a given contour signal, the 1-D Fourier
center point. transform is given as;


A histogram containing each bin of the shape context 1XN‐1
‐j2πnt
FDn ¼ sðtÞ  exp ð3Þ
(SCH) is constructed for each part of the shape context N t¼0 N
to enable the shape context to viable to various transfor-
mations. In the wake of the fact that the Fourier Trans- where:
form is widely used transformations (Ghazal et al. 2009;
Zhang and Lu 2005) for object recognition problems s(t) represents the 1-D contour signal,
and its coefficients are found to be invariant to sym- N represents number of representative points of the
metry operations (i.e. translation, rotation and scaling contour,
etc). The size of the shape representation points is an n = 0,1,2,…,N-1 and,
important and influential factor that optimizes the utility FDn represents nth Fourier descriptor.
of Fourier transformation. Hence, in the shape signature
generation process, the sampling is considered as a Using Equation (3), the required Fourier Descriptors
mandatory step. Some of the sampling methods like of size ‘N’ are generated. Further, the extracted features
Equal Point (EP), Equal Angle (EA) and Equal Arc are testified for their invariance to translation, rotation
Length (EAL) (Zhang and Lu 2003) are considered pres- and scaling operations (performed over the set of im-
ently. EAL is expected to yield for a better equal space ages). In the wake of the fact that the proposed descrip-
(Peter and Otterloo 1991) than the other two methods. tor is obtained with respect to the centroid, the obtained

Figure 2 Example of constructing (a) Belongie Shape Context (b) Modified Shape Context signature for a contour point wrt the
corresponding farthest point.
Madireddy et al. SpringerPlus 2014, 3:674 Page 4 of 12
http://www.springerplus.com/content/3/1/674

Figure 3 Proposed descriptor for three Animal group Images (a1) Animal 3 (a2) MSC signature of Animal 3 (b1) Animal 5 (b2) MSC
signature of Animal 5 (c1) Animal 7 (c2) MSC signature of Animal 7.

features are expected to be invariant to translation. The this, the average global distance of global feature vec-
possible finite (and stipulated) magnitude of the values tors is directly added to the Euclidean distance of the
for the features vouches for the rotation invariance. For Fourier descriptor feature vector.
the present method, the scaling invariance is also pre- The distance between two shape context vectors in-
sented by involving the process of dividing the features cluding the object global feature vectors is given by the
with the first feature value. In the third step, the feature Equation (5).
vector is constructed, which describes the entire shape
features of the object. DX ðTE; TRÞ
DðTE; TRÞ ¼ EDðTE; TRÞ þ ð4Þ
To further improve the quality of proposed methods, 3
the global information of the object is also considered.
For this, experiments are conducted with considering where:
different global descriptors and identified that three glo-
bal descriptors (GD) are efficient to represent the global ED (TE, TR) represents the Euclidean Distance
shape information. The GD feature vector, viz., {S, C, A} between the test and trained shapes and,
contains the measures of solidity, circularity and aspect DX (TE,TR) represents the Global distance between the
ratio is computed for the given object. test and trained shapes.
In the fourth step, the shape toning process is executed.  
X XTE ‐XTR 
In the shape toning process, the distance measures DX ðTE; TRÞ ¼  TR  ð5Þ
(Ghazal et al. 2009, 2012) used is viz., the Euclidean dis- X max X
tance (ED). The distance measure between two objects
shape context vectors is given by the Equation (4). In where:

Figure 4 Proposed descriptor for three Hand group Images (a1) Hand 8 (a2) MSC signature of Hand 8 (b1) Hand 9 (b2) MSC signature
of Hand 9 (c1) Hand 11 (c2) MSC signature of Hand 11.
Madireddy et al. SpringerPlus 2014, 3:674 Page 5 of 12
http://www.springerplus.com/content/3/1/674

Figure 5 Proposed descriptor for three Heart group Images (a1) Heart 7 (a2) MSC signature of Heart 7 (b1) Heart 11 (b2) MSC
signature of Heart 11 (c1) Heart 12 (c2) MSC signature of heart 12.

X represents the GD vector {S, C, AR}, Performance

XTE represents the GD feature of the test shape and, The performance of various object recognition schemes
XTR represents the GD feature of the trained shape. reported (Shu and Xiao Jun 2011; Ghazal et al. 2009,
2012) so far employ different measures. Among them,
According to the specificity of the data of distance precision and recall are considered as important mea-
measurement, the distances are further rearranged in as- sures, while they verbally quantify the similarity
cending order and are assigned with ranks. However, the measurement. Precision (P) and Recall (R) are defined
system is also enabled to recognize and register the top by:
ranked images. x
The standard databases (Sikora 2001; Sebastian et al. P¼ ð6Þ
y
2001) used for the evaluation of shape descriptors
presently are Kimia {K-99, K-216} and MPEG CE-1 Set where:
B. It is noticed that the Set B database with 70 groups
and each group with 20 images. It characteristically x denotes the true recognition results,
includes rotated, scaled, skewed and defected shapes. y denotes the total recognized result and,
However the K-99 database which consists of 9 groups, P denotes the precision.
each group with 11 images. It is known to include the
partially occluded shapes. The K-216 database with 18 x
groups, each group with 12 images, it represents a sub R¼ ð7Þ
groupsize
database of Set B, and contains partially occluded
shapes. where:

Figure 6 Proposed descriptor for three Fly group Images (a1) Fly 1 (a2) MSC signature of Fly 1 (b1) Fly 4 (b2) MSC signature of Fly 4
(c1) Fly 10 (c2) MSC signature of Fly 10.
Madireddy et al. SpringerPlus 2014, 3:674 Page 6 of 12
http://www.springerplus.com/content/3/1/674

Figure 7 Retrieval results of Key2 test image from K-99 database (a) Original image (b) ZMD result (c) BSC+GD result (d) MSC+GD result.

R denotes the Recall, FDR denotes the False Detection Rate,

x denotes the true recognition results and, z denotes the false recognition result and,
group size denotes the maximum true recognition y denote the total recognized result.
result.
The average FDR (AFDR) value of all test images cor-
The Average Precision value for each recall is com- responding to each database is estimated. Apart from
puted. This value is affirmatively grouped as two cat- the usual recognition rate, the Average Processing Time
egories viz., Low Recall (LR), High Recall (HR). The (APT) is also estimated for each query in the shape ton-
Average Precision for Low Recall (APLR) denotes the ing stage. The proposed descriptor is compared with 4
average precision for recalls less than or equal to 50. In standard descriptors viz., Angular Radial Transform
contrast, the Average Precision for High Recall (APHR) Descriptor (ARTD) (Zhang et al. 2008), Moment In-
represents the average precision for recalls greater than variant Descriptor (MID) (Zhang et al. 2008), Zernike
50. The False Detection Rate (FDR) for each of the Moment Descriptor (ZMD) (Tiagrajah and Razeen
image is also estimated by: 2011) and Curvature-Scale-Space-Descriptor (CSSD)
(Tiagrajah and Razeen 2011). A specific feature size of
z 35 for ARTD (n < 3, m < 12), 6 for MID, 34 for ZMD
FDR ¼ ð8Þ (order from 2 to 10) is adopted. The CSSD feature size
y
is varying from that for one image to another image
where: since number of peaks is varying. All the cited metrics viz.,

Figure 8 Retrieval results of Fork12 test image from K-216 database (a) Original image (b) ZMD result (c) BSC+GD result
(d) MSC+GD result.
Madireddy et al. SpringerPlus 2014, 3:674 Page 7 of 12
http://www.springerplus.com/content/3/1/674

Figure 9 Retrieval results of Carriage16 test Image from Set B database (a) Original image (b) ZMD result (c) BSC+GD result
(d) MSC+GD result.

APLR, APHR, AFDR and APT are evaluated to estimate K-216 databases. Figures 3(a1), -3(b1) and -3(c1) contains
the performance for the proposed descriptor (with inclu- three original images of animal group (animal-3, animal-5
sion of GD), ARTD, MID, ZMD and CSSD. and animal-7); the Figures 4(a1), -(b1) and -(c1) pertain to
the three original images of hand group (hand8, hand9,
Results and discussion hand11); the Figures 5(a1), -(b1) and -(c1) give three
Shape context based object recognition is estimated as original images of heart group (heart-7, heart-11 and
detailed in section-Design of system for the input of heart-12); and Figures 6(a1), -(b1) and -(c1) contain three
standard databases. The trends of the results that fol- original images of fly group (Fly1, Fly4 and Fly10). The
low various approaches are presented in the following corresponding Modified shape context (MSC) histogram
sub section-Processing of modified shape context based is illustrated in Figures 3(a2), 4, 5, 6(a2), Figures 3(b2), 4,
object recognition. The relative performance of the 5, 6(b2) and Figures 3(c2), 4, 5, 6(c2) respectively. It is
proposed descriptor is also analyzed in the section- clearly noticed from Figures 3, 4, 5, 6, that the MSC
Performance evaluation in the wake of the other re- histogram is found to be similar for the different shapes
ported methods. within the same group; while it exhibits difference be-
tween those of one group compared to the other. Basing
Processing of modified shape context based object on 1-D shape signal, the Fourier descriptors (FD) are
recognition generated. The experiments are conducted with varying
The shape context is constructed with 60 bins. Then for size of feature vector. From this, it is identified that the
each contour point, the angle is measured (i.e. BSC and first ten features of the FD are consistent. Hence, they
MSC) within the range of one full rotation i.e. from 0° to are used to design the feature vector.
360°. A histogram is generated that corresponds to each In the present object recognition process, Euclidean
and every bin of shape context. The histograms, thus Distance (ED) measure of performance is estimated
constructed are presented in Figures 3, 4, 5, and 6 corre- between the target and test objects, while they are al-
sponding to four different image groups (i.e. animal, located with ranking according to their distance. In ac-
hand, heart and glass) as accessed from set B, K-99 and cord with the established procedures, the top n-ranked

Table 1 The APLR and APHR values for various Table 2 The APLR and APHR values for various
descriptors with Set B database descriptors with K-99 database
Avg. precision Avg. precision
Descriptor Recall < =50% Recall >50% Average Descriptor Recall < =50% Recall >50% Average
BSC 82.65 49.37 66.01 BSC 86.01 51.25 68.63
MSC 84.02 49.56 66.79 MSC 89.11 56.87 72.99
BSC + GD 84.48 52.85 68.67 BSC + GD 87.21 57.82 72.52
MSC + GD 88.83 55.58 72.21 MSC + GD 90.77 62.43 76.60
ARTD 82.10 45.69 63.90 ARTD 84.26 45.72 64.99
MI 79.54 44.50 62.02 MI 81.96 44.74 63.35
ZMD 82.56 45.62 64.09 ZMD 89.61 61.37 75.49
CSSD 78.61 41.81 60.21 CSSD 82.32 44.11 63.22
Madireddy et al. SpringerPlus 2014, 3:674 Page 8 of 12
http://www.springerplus.com/content/3/1/674

Table 3 The APLR and APHR values for various Figure 8(c) gives the retrieval results with BSC + GD de-
descriptors with K-216 database scriptor, and Figure 8(d) gives the retrieval results with
Avg. precision MSC + GD descriptor.
Descriptor Recall < =50% Recall >50% Average For Set B database:- the top 20 ranked images corre-
BSC 88.50 57.59 73.05 sponds to the query image of Carriage16 by using ZMD,
MSC 90.16 61.18 75.67
BSC + GD and MSC + GD as presented in the Figures 9
(a)-(d). The Figure 9(a) gives the Carriage16 query
BSC + GD 91.01 64.00 77.50
image, Figure 9(b) gives the retrieval results with ZMD
MSC + GD 91.50 66.00 78.75 descriptor, Figure 9(c) gives the retrieval results with
ARTD 81.35 44.67 63.01 BSC + GD descriptor and Figure 9(d) gives the retrieval
MI 80.14 46.04 63.09 results with MSC + GD descriptor. Overview of cited
ZMD 88.94 61.71 75.33 Figures 7(a), 8, 9(d) suggests that, the MSC + GD de-
CSSD 80.12 44.97 62.55
scriptor performs better for retrieval of more relevant
images with relatively strong correspondence than that
with the other descriptors.
objects are used to estimate the Precision and Recall pa-
rameters. The ‘n’ notifies the group size for {Set B:20, K- Performance evaluation
99:11 and K-216:12} sets. For each database, the accur- The yield of APLR and APHR values for the descriptor
acy for the retrieval results corresponding to top ‘n’ with the currently proposed distance measure is ana-
(group size of the corresponding database) number of lyzed. In the wake of the four other standard descrip-
images is estimated and illustrated in Figures 7, 8, 9 re- tors, the aspect of compatibility (with three databases)
spectively. For K-99 database:- The top 11 ranked images is also explored, while the results are presented in
correspond to the query image Key2 (as estimated by Tables 1, 2, 3 respectively. From these results, it is
using ZMD, BSC + GD and MSC + GD), while they are clearly evident that the presently proposed descriptor
presented in the Figures 7(a)-(d). The Figure 7(a) corre- out performs the other descriptors regarding all the
sponds to the Key2 query image and Figure 7(b) corre- three standard databases. However, among the pres-
sponds to the retrieval results with ZMD descriptor. ently considered descriptors, the CSSD descriptor is
Figure 7(c) gives the retrieval results for BSC + GD de- found to accompany with a lower performance, fol-
scriptor; and Figure 7(d) vouches for the retrieval results lowed by that of MID, ARTD. However, the case of
with MSC + GD descriptor. For K-216 database:- The ZMD resulted for the next higher performance. But, for
top 12 ranked images corresponding to the query image Set B database, the ZMD is yielding the highest result.
Fork12 are estimated by using ZMD, BSC + GD and From Table 1, it is found that the proposed MSC + GD
MSC + GD as presented in Figures 8(a)-(d). The Figure 8(a) happen to be influential to increase the APLR value of
illustrates the Fork12 query image and Figure 8(b) ZMD. It is also found to significantly increase the
speaks for the retrieval results with ZMD descriptor. APHR value of ARTD. For K-99 and K-216 databases,

100

90

80

70

60
Precision

50 MSC+GD
ARTD
40
MID
30
ZMD
20 CSSD
10

0
0 10 20 30 40 50 60 70 80 90 100
Recall
Figure 10 The PR graph for various descriptors with set B database.
Madireddy et al. SpringerPlus 2014, 3:674 Page 9 of 12
http://www.springerplus.com/content/3/1/674

100

90

80

70

60
Precision

50 MSC+GD

40 ARTD

MID
30
ZMD
20
CSSD
10

0
0 10 20 30 40 50 60 70 80 90 100
Recall
Figure 11 The PR graph for various descriptors with K-99 database.

the ZMD is giving distinctly improved APLR and APHR database is depicted in Figure 11. From this figure, it is
values than with the other descriptors. From the observed that the ZMD outperforms other standard de-
Tables 2 and 3; it is evident that the proposed MSC + GD scriptors with increased precision measure in the range
is accompanied with an improved performance in terms of of both lower and higher recalls. An overview of the
enhancement of APLR and APHR. results infers that the proposed MSC + GD measure is
The PR plots for these five descriptors comparing to considerable increase in precision measure at lower recalls
the set B, K-99 and K-216 are presented in Figures 10, ranged between 40 and 50 and higher recalls ranged be-
11 and 12 respectively. Figure 10 reveals that all the five tween 60 to 70 and 90 to 100. The precision measure is
descriptors are yielding considerable enhanced perform- found to attain considerable improvement at higher recalls
ance with regard to the precision measure for the Set B i.e. at 80 to 100. Figure 12 describes the PR plot of various
database in the range of low recalls. At higher recalls, descriptors for K-216 database. The ZMD in this PR plot
the proposed MSC + GD measure is found to result for is found to be superior, rather than other standard de-
improved precision measure, rather than the other scriptors at lower and higher recalls. Thus the proposed
standard descriptors. The proposed CSD is found to in- MSC + GD measure is giving increased precision measure
crease the precision measure marginally at lower recalls regarding PR plots at both lower and higher recalls.
i.e. ≤50, bit, it is found to significantly increase the preci- Other Performance measures viz., Average False Disco-
sion at higher recalls i.e. >50. The PR plot for K-99 very Rate (AFDR) and Average Processing Time (APT) are

100

90
80
70
Precision

60

50
MSC+GD
40
ARTD
30
MID
20 ZMD
10 CSSD
0
0 10 20 30 40 50 60 70 80 90 100
Recall
Figure 12 The PR graph for various descriptors with K-216 database.
Madireddy et al. SpringerPlus 2014, 3:674 Page 10 of 12
http://www.springerplus.com/content/3/1/674

Table 4 The AFDR for various descriptors with three Table 6 Bull’s eye score for various descriptors with set B
databases database
Descriptor AFDR Descriptor Score %
Set B K-99 K-216 CSS (Mokhtarian et al. 1997) 75.44
MSC + GD 0.70 0.72 0.73 Visual Parts (Latecki and Rolf 2000) 76.45
ARTD 0.90 0.85 0.86 SC + TPS (Belongie et al. 2002) 76.51
MID 0.91 0.81 0.84 Aligning Curves (Sebastian et al. 2003) 78.16
ZMD 0.78 0.76 0.76 SSC (Xie et al. 2008) 79.92
CSSD 0.84 0.88 0.89 CPDH + EMD (Cem Direko glu and Nixon 2011) 76.56
General Model (Tu and Yuille 2004) 80.03
also estimated as detailed in section-2.2, while the esti- MSC + GD 81.64
mated AFDR values for the three databases are pre-
IDSC + DP (Ling & Jacobs 2007) 85.40
sented in Table 4. It is observed that the proposed
MSC + GD results for a lower value for all the three da-
tabases. Since, the APT measure is argued to exhibit  The proposed descriptor improves the precision
profound influence on shape toning stage (for each of measures at high recalls when compared with the
the shape descriptor), it is also estimated for all the low recalls thus enabling more relevant objects to be
three databases, and presented in Table 5. The proposed recognized.
MSC + GD is found to yield for less APT value in com-  With less feature vector size, the proposed
parison with other descriptors. Therefore, basing the descriptor enables the object recognition system to
observed trends of performance measures, it is argues be efficient with less APT and AFDR measures.
that the proposed descriptor exhibits higher efficiency.
As a measure of performance, the retrieval rate with Competing interests
PSVG areas of interest include Object Oriented Technologies, Information
bull’s eye score (Cem Direko glu and Nixon 2011) is
Retrieval, Algorithms, Computer Networks, and Image Processing. PDM
also estimated. This measure involves the calculation of areas of interests are Experimental soft condensed matter, Liquid crystals-
the ratio of the total number of shapes (i.e. from the design and characterization, Theory of spectroscopy, chemistry – design of
supermolecules, Face Recognition, Image Processing, and optical character
same class) to the highest possible number of shapes in
recognition. SCH area of interest includes Image processing, Database
the same database. The estimated bull’s eye score for Management Systems, Speech Recognition, Pattern recognition and
top 40 results in Set B database is presented in Table 6. network security.
It is clear that Inner Distance Shape Context (IDSC) is
yielding highest score when compared with the others. Authors’ contributions
RMM carried out the shape context studies, participated in the
However, as this includes complicated dynamic pro- development of the proposed shape descriptor and drafted the
gramming procedure, the simple Euclidean distance manuscript. PSVG participated in the design of the study with the shape
measure for the proposed descriptor is argued to be toning process, and drafted the manuscript. PDM participated in the
design of the proposed shape descriptor with Fourier Transformation, and
more efficient retrieval parameter. drafted the manuscript. SCH participated in the design of the study with
feature extraction, and drafted the manuscript. All authors read and
Conclusions approved the final manuscript.

Author details
 Shape context based description is proved to be 1
Department of CSE, Pragati Engineering College, Surampalem, AP 533437,
efficient when compared with various other India. 2Department of IT, SRKR Engineering College, Bhimavaram, AP 534204,
India. 3Department of Physics, Jawaharlal Nehru Technological University
standard descriptors with respect to various Kakinada, Kakinada, AP 533003, India. 4Department of CSE, Jawaharlal Nehru
performance measures viz., APLR, APHR, AFDR Technological University Kakinada, Kakinada, AP 533003, India.
and APT.
Received: 10 September 2014 Accepted: 31 October 2014
Published: 15 November 2014

Table 5 APT of various descriptors with set B database

References
Descriptor APT Abbasi S, Mokhtarian F, Kittler J (1999) Curvature scale space image in shape
MSC + GD 0.0011 similarity retrieval. Multimed Syst 7:467–476
Abbasi S, Mokhtarian F, Kittler J (2000) Enhancing CSS-based shape retrieval for
ARTD 0.0017 objects with shallow concavities. Image Vis Comput 18:199–211
MID 0.0302 Adamek T, O’Connor NE (2004) A multi scale representation method for non
rigid shapes with a single closed contour. IEEE Trans Circuits Syst Video
ZMD 0.0017 Technol 14:742–753
Alajlan N, Rube IE, Kamel MS, Freeman G (2007) Shape retrieval using triangle-area
CSSD 2.1640
representation and dynamic space warping. Pattern Recogn 40:1911–1920
Madireddy et al. SpringerPlus 2014, 3:674 Page 11 of 12
http://www.springerplus.com/content/3/1/674

Arica N, Vural FTY (2003) BAS: a perceptual shade descriptor based on the beam Mokhtarian F, Mackworth A (1986) Scale-based description and recognition of planar
angle statistics. Pattern Recogn Lett 24:1627–1639 curves and two-dimensional shapes. IEEE Trans Pattern Anal Mach Intell 8:34–43
Attalla E, Siy P (2005) Robust shape similarity retrieval based on contour Mokhtarian F, Abbasi F, Kittler J, Smeulders AWM, Jain R (1997) Efficient and
segmentation polygonal multi resolution and elastic matching. Pattern robust retrieval by shape content through curvature scale space. In: Image
Recogn 38:2229–2241 Databases and Multi-Media Search. World Scientific Publishing, Singapore,
Bartolini I, Ciaccia P, Patella M (2005) WARP: accurate retrieval of shapes using pp 51–58
phase of Fourier descriptors and time warping distance. IEEE Trans Pattern Mori G, Malik J (2003) Recognizing objects in adversarial clutter: breaking a visual
Anal Mach Intell 27:142–147 CAPTCHA. IEEE Conf Comput Vision Pattern Recogn 1:1063–6919
Belkasim SO, Shridhar M, Ahmadi M (1991) Pattern recognition with moment Mukundan R, Ong SH, Lee PA (2001) Image analysis by Tchebichef moments.
invariants: a comparative study and new results. Pattern Recogn IEEE Trans Image Process 10:1357–1364
24:1117–1138 Nixon MS, Aguado AS (2002) Feature Extraction and Image Processing, 1st edn.
Belongie S, Malik J, Puzicha J (2002) Shape matching and object Newnes Publishers, Burlington MA, pp 247–287
recognition using shape contexts. IEEE Trans Pattern Anal Mach Intell Peter J, Otterloo V (1991) A contour- Oriented Approach to Shape Analysis.
24(4):509–522 Prentice Hall, Hertfordshire UK
Bookstein FL (1989) Principal warps: thin-plate-splines and decomposition of Petrakis EGM, Diplaros A, Milios E (2002) Matching and retrieval of distorted and
deformations. IEEE Trans Pattern Anal Mach Intell 11(6):567–585 occluded shapes using dynamic programming. IEEE Trans Pattern Anal Mach
Bronstein MM, Bronstein AM (2011) Shape recognition with spectral distances. Intell 24:1501–1516
IEEE Trans Pattern Anal Mach Intell 33(5):1065–1071 Salve SG, Jondhale KC (2010) Shape matching and object recognition using
Bronstein MM, Kokkinos I (2010) Scale invariant heat kernel signatures for non shape contexts. In: Computer Science and Information Technology (ICCSIT),
rigid shape recognition. IEEE Conf Comput Vision Pattern Recogn (CVPR) vol 9. 3rd IEEE International Conference, pp 471–474
1704–1711, doi:10.1109/CVPR.2010.5539838 Sebastian TB, Klein PN, Kimia BB (2001) Recognition of shapes by editing shock
Cem Direko glu, Nixon MS (2011) Shape classification via image-based multi scale graphs. ICCV 1:755–762
description. Pattern Recogn 44:2134–2146 Sebastian T, Klein P, Kimia B (2003) On aligning curves. IEEE Trans Pattern Anal
Chauang G, Kuo C (1996) Wavelet descriptor of planar curves: theory and Mach Intell 25(1):116–125
applications. IEEE Trans Image Process 5:56–70 Sebastian TB, Klein PN, Kimia BB (2004) Recognition of shapes by editing their
Daliri MR, Torre V (2008) Robust symbolic representation for shape recognition shock graphs. IEEE Trans Pattern Anal Mach Intell 26(5):550–571
and retrieval. Pattern Recogn 41:1782–1798 Shu X, Xiao Jun W (2011) A novel contour descriptor for 2D shape
Day AM, Arnold DB, Havemann S, Fellner DW (2004) Combining polygonal and matching and its application to image retrieval. Image Vis Comput
subdivision surface approaches to modeling and rendering of urban 29:286–294
environments. Comput Graph 28(4):497–507 Siddiqi K, Shokoufandeh A, Dickinson SJ, Zucker SW (1999) Shock graphs and
Dubois SR, Glanz FH (1986) An autoregressive model approach to two- dimensional shape matching. Int J Comput Vis 35(1):13–32
shape classification. IEEE Trans Pattern Anal Mach Intell 8:55–65 Sikora T (2001) The MPEG-7 visual standard for content description- an overview.
Flusser J (2000) On the independence of rotation moment invariants. Pattern IEEE Trans Circuits Syst Video Technol 11(6):696–702
Recogn 33:1405–1410 Tao Y, Wi G (1999) Delaunay Triangulation for Image Object Indexing: A
Forsyth D, Mundy J (1999) Shape, contour and grouping in computer vision. Lect Novel Method for Shape Representation. Seventh SPIE Symposium on
Notes Comput Sci 1681:1–3 Storage and Retrieval for Image and Video Databases San Jose, CA,
Ghazal AE, Basir O, Belkasim S (2009) Farthest point distance: a new shape pp 631–642
signature for Fourier descriptors. Signal Process Image Comm Teague M (1980) Image analysis via the general theory of moments. J Opt Soc
24:572–586 Am 70:920–930
Ghazal AE, Basir O, Belkasim S (2012) Invariant curvature-based Fourier shape Thayananthan A, Stenger B, Torr PHS, Cipolla R (2003) Shape context and
descriptors. J Vis Commun Image R 23:622–633 chamfer matching in cluttered scenes. IEEE Conf Comput Vision Pattern
Goshtasby A (1985) Description and discrimination of planar shapes using shape Recogn 1:127–133
matrices. IEEE Trans Pattern Anal Mach Intell 7:738–743 Tiagrajah VJ, Razeen AASM (2011) An enhanced shape descriptor based on radial
Grigorescu C, Petkov N (2003) Distance sets for shape filters and shape distances. IEEE Int Conf Signal Image Process Appl (ICSIPA) 472–477,
recognition. IEEE Trans Image Process 12(10):1274–1286 doi:10.1109/ICSIPA.2011.6144073
Hu M (1962) Visual pattern recognition by moment invariants. IRE Trans Inform Tu Z, Yuille A (2004) Shape matching and recognition-using generative models
Theor IT 8:115–147 and informative features. Proc Eur Conf Comput Vis 3:195–209
Iyer N, Jayanti S, Lou K, Kalyanaraman Y, Ramani K (2005) Three dimensional Wallace TP, Wintz PA (1980) An efficient three dimensional aircraft recognition
shape searching: state of the art review and future trends. Comput Aided algorithm using normalized Fourier descriptors. Comput Graph Image
Des 37:509–530 Process 13:99–126
Junding S, Xiaosheng W (2006) Chain Code Distribution-Based Image Retrieval. Xie J, Heng P, Shah M (2008) Shape matching and modeling using skeletal
International Conference on Intelligent Information Hiding and Multimedia context. Pattern Recogn 41(5):1756–1767
Signal Processing, China, pp 139–142 Xingyuan W, Zongyu W (2013) A novel method for image retrieval based on
Khotanzad A (1990) Invariant image recognition by Zernike moments. IEEE Trans structure elements’ descriptor. J Vis Commun Image R 24:63–74
Pattern Anal Mach Intell 12:489–497 Yadav RB, Nishchal NK, Gupta AK, Rastogi VK (2007) Retrieval and classification
Konukoglu E, Glocker B, Criminisi A, Pohl KM (2013) WESD-weighted spectral of shape-based objects using Fourier, generic Fourier, and wavelet-
distance for measuring shape dissimilarity. IEEE Trans Pattern Anal Mach Intell Fourier descriptors technique: a comparative study. Opt Lasers Eng
35(9):2284–2297 45:695–708
Kunttu I, Lepisto L, Rauhamaa J, Visa A (2006) Multi scale fourier descriptors for Yang M, Kpalma K, Ronsin J (2008) A Survey of shape feature extraction
defect image retrieval. Pattern Recogn Lett 27:123–132 techniques. In: Pen Y (ed) Pattern Recognition., pp 43–90
Latecki LJ, Lakamper R (2000) Shape similarity measure based on correspondence Zahn T, Roskies RZ (1972) Fourier descriptors for plane closed curves. IEEE Trans
of visual parts. IEEE Trans Pattern Anal Mach Intell 22(10):1185–1190 Comput 21:269–281
Lateckia LJ, Lakaempera R, Wolter D (2005) Optimal partial shape similarity. Image Zhang D, Lu G (2002) Shape-based image retrieval using generic Fourier descriptor.
Vis Comput 23:227–236 Signal Process Image Commun 17:825–848
Li S, Lee MC (2005) Effective invariant features for shape-based image retrieval. Zhang D, Lu G (2003) A comparative study of curvature scale space and Fourier
J Am Soc Inf Sci Technol 56:729–740 descriptors for shape based image retrieval. J Vis Commun Image Represents
Ling H, Jacobs DW (2007) Shape classification using the inner distance. IEEE Trans 14(1):39–57
Pattern Anal Mach Intell 29(2):286–299 Zhang D, Lu G (2004) Review of shape representation and description
Lu G, Sajjanhar A (1999) Region-based shape representation and similarity techniques. Pattern Recogn 37(1):1–19
measure suitable for content based image retrieval. Multimed Syst Zhang D, Lu G (2005) Study and evaluation of different fourier methods for
7:165–174 image retrieval. Image Vis Comput 23(11):33–49
Madireddy et al. SpringerPlus 2014, 3:674 Page 12 of 12
http://www.springerplus.com/content/3/1/674

Zhang H, Malik J (2003) Learning a discriminative classifier using shape context

distances. IEEE Conf on Comput Visi Pattern Recogn 1:242–247
Zhang G, Ma ZM, Tong Q, He Y, Zhao T (2008) Shape Feature Extraction using
Fourier Descriptors with Brightness in Content-Based Medical Image
Retrieval. International Conference on Intelligent Information
Hiding and Multimedia Signal Processing, pp 71–74

doi:10.1186/2193-1801-3-674
Cite this article as: Madireddy et al.: A modified shape context method
for shape based object retrieval. SpringerPlus 2014 3:674.

Submit your manuscript to a

journal and benefit from:
7 Convenient online submission
7 Rigorous peer review
7 Immediate publication on acceptance
7 Open access: articles freely available online
7 High visibility within the field
7 Retaining the copyright to your article

Submit your next manuscript at 7 springeropen.com