Optimised Kd-Tree Approach with Dimension Reduction for Efficient Indexing and Retrieval from Multibiometric Database

Optimised Kd-Tree Approach with Dimension
Reduction for Efficient Indexing and Retrieval
from Multibiometric Database
Revathi B
Department of ECE
Pondicherry Engineering College
Puducherry, India
brevathi@pec.edu
Gnanakumar D
Department of ECE
National Institute of Technology
Trichirappalli, India
gk.wc@pec.edu
Dr. Gnanou Florence Sudha
Professor, Department of ECE
Pondicherry Engineering College
Puducherry, India
gfsudha@pec.edu
Abstract— Biometric retrieval is a challenging task as the size of
the databases have increased considerably. In this work, a novel
optimized kd-tree algorithm is implemented to enhance the
efficiency of indexing and retrieving for a multibiometric
database comprising of iris and fingerprints. To improve the
retrieval performance, fingerprint image is represented by
minutiae features and iris image is represented by texture
features and the features are fused together by feature level
fusion. Dimension reduction of the feature vector is carried out
using Principal Component Analysis to reduce the storage space
and increase retrieval rate. The proposed optimized kd-tree
indexing technique with dimension reduction aims to overcome
the limitations of the existing nearest kd-tree. From the
experimental results, it is concluded that the proposed optimized
kd-tree indexing algorithm with dimension reduction has reduced
False Acceptance Rate and False Rejection Rate and has
improved Hit rate to 95% at 60% penetration rate compared to
existing nearest kd-tree techniquefora multibiometricdatabase.
Keywords- Indexing; Multibiometrics; Nearestkd-tree indexing;
Optimised kd-tree indexing; PCA;.
I. INTRODUCTION
Biometrics is the science of uniquely recognizing humans
based upon one or more intrinsic physical or behavioral traits.
The number of biometric systems deployed on an international
scale is increasing and the size of the underlying databases
also increases manifold. As these databases are used
extensively for many applications, efficient ways for searching
and retrieving relevant identities becomes imperative.
Indexing is a process of providing a means for uniquely
classifying data which in turn makes items easier to retrieve.
Several Indexing techniques for biometric databases have been
proposed. But existing systems have drawbacks such as search
time and accuracy [1-3]. Compared to other methods, indexing
using multibiometric is an area which is rapidly growing
because of its great benefits like accuracy, reliability and
provides relatively good performance [4].Fingerprints and iris
are the most widely used parameters for personal identification
amongst all biometrics. Most of the finger-scan technologies
are based on Minutiae [5], while the eye-scan technologies are
based on Iris [6]. Minutia-based techniques represent the
fingerprint by its local features, like terminations and
bifurcations and Iris-based techniques represent the eye by its
texture features that are unique complex random patterns. This
approach has been intensively studied, and is also the
backbone of the current available multibiometric recognition
products [7-8].
Many biometric feature techniques have been proposed in
the literature but still there is need for more accurate, fast and
efficient algorithms [9]. Many of the existing techniques use
single identity or multiple identity features [19]. These
features are computationally expensive to compute and there is
no standard definition of these features since these features are
not defined in the standard template [10-11]. In addition,
template size has to be kept as minimum as possible so that the
algorithm works efficiently. It has been suggested [12-13] that
if the statistical feature of fused biometric information is used
for the purpose of multibiometric indexing, it will be more
tolerant and robust against global transformations and non-
linear distortions. It will also give better results than using
global information [14]. Hence a statistical feature extracted
from the fused template has been used in this work.
In this paper, an efficient indexing technique which can be
used in an identification system with large biometric database
has been proposed. To make the indexing algorithm fast and
robust, attention is focused on a good indexing technique. This
International Journal of Computer Science and Information Security (IJCSIS),
Vol. 15, No. 8, August 2017
224 https://sites.google.com/site/ijcsis/
ISSN 1947-5500

paper proposes to exploit the merits of the optimisedkd-tree
indexing technique [15] coupled with high dimensional index
selectivity for indexing the multibiometric database. The
paper focuses on the choice of statistical index features, such
that it is invariant to distortion. Dimension reduction of the
fused feature vector is carried out to reduce the storage space
for large databases and reduce the time required for matching.
To show the improvements of the proposed optimized kd-tree
indexing technique, existing nearest kd-tree indexing is
considered as the reference technique. The performance of the
proposed technique is analyzed using MATLAB simulation
and various performance metrics are computed to demonstrate
its superior performance.
The paper is organized as follows. The present section gives
the need for efficient indexing technique and related works. In
Section 2, the existing nearest kd-tree indexing and proposed
optimized kd-tree indexing are discussed. Section 3 explains
the dimension reduction technique and its implementation.
The simulation results are discussed in Section 4. Finally
conclusion is made in Section 5.
II. PROPOSED SYSTEM MODEL
Traditional databases are indexed numerically or
alphabetically to reduce the retrieval time of records from the
database. In biometric databases, biometric templates do not
possess any natural or alphabetical order. Hence, there is a
need for an efficient indexing technique which stores the
Fig 1. Enrollment stage
Fig. 2 Recognition stage
templates in some predefined manner so that an efficient
retrieval strategy can be achieved. In this work, a
multibiometric database comprising of fingerprint and iris has
been used.
The proposed biometric recognition system consists of two
subsystems; one for enrollment and second one for
recognition. In the enrollment stage, the biometrics namely,
fingerprint and iris are acquired from the individuals, feature
sets are extracted from the acquired data and the features are
fused together by feature level fusion. From the fused feature
set, the statistical feature f(x,y) such as mean and standard
deviation are computed and stored in the database as shown in
figure 1(a).
Statistical features are more tolerant and robust against
global transformations and non-linear distortions and are
hence used in this work. Multiple templates per individual are
typically used to compute the statistical features for accuracy
considerations.
In the recognition stage, biometric data, i.e, fingerprint and
iris of the individual is captured, a feature set is computed
(query template) and then it is compared with the templates
stored during enrollment in the database, as shown in figure
1(b). Thus biometric acquisition, preprocessing and feature
extraction stages are required to create a template. To reduce
storage space and increase retrieval rate, the reduced
dimensional feature vectors are fused at feature level and the
fused feature vectors are used to index the database by
forming optimized Kd-tree for efficient retrieval. During the
retrieval phase, the query template of the user is compared
with other dimension reduced indexed templates, estimating
the similarity between them. Based on the matching result, the
user is authenticated.
A. Nearest kd tree indexing algorithm
kd-tree is a binary tree that represents hierarchical
subdivision of space using splitting planes as shown in Figure
2. It is a space partitioning data structure for organizing points
in a k dimensional space and useful for searches involving
multi-dimensional search key. k denotes the dimensionality of
the template and the number of nodes in a kd-tree is equal to
the number of templates in the gallery set. Each node in a kd-
tree consists of information field, discriminator field, value
field and two pointer fields. Information field contains
descriptive information about the node, value field contains
feature vector values, discriminator field takes a value from 1
to k and pointer fields has pointer value of left and right
subtree of the node. At the root, data points are split into two
halves by a partition hyper plane and each half is assigned to
one child node. Each of the two halves is recursively split in
the same way to create a balanced binary tree.
To find the nearest neighbor for a query template, a top
down searching procedure is performed from root of kd- tree
to a leaf node. According to the nearest neighbor problem,
given a point set P of n objects, search the nearest point in P
for given query point q.
Vol. 15, No. 8, Augus 2017
ISSN 1947-5500

This problem can be called as nearest kd-tree search, if k
points in P which are nearest to q are required as the result. It
is called as R-near neighbors, if all points within distance R
Fig. 2 Pictorial representation of kd-tree structure
from point q are required as the result. In approximation based
nearest search algorithm, a point is returned as result if, its
distance from the query is at most c times the distance from
the query to its nearest neighbor where c is the approximation
factor. This is known as c-approximate nearest neighbor’s
problem. This can also be generalized to c-approximate kd-
tree closest point neighbor problem and c-approximate R-
nearest neighbor problem.
The kd-tree algorithm has better performance compared to
other indexing techniques such as k-nearest neighbour
algorithm [16]. However, the drawback of nearest kd-tree
algorithm is that it performs poorly if the training set is large.
In kd-trees, the partition hyper plane is limited to be
orthogonal to the coordinate axis. This results in the space not
being well partitioned. Hence more leaf nodes are to be
searched to get the nearest neighbors instead of considering
only one node.
(a) Nearest neighbors on kd-tree
(b) Left subtree on nearest kd-tree
In addition, it is very sensitive to irrelevant or redundant
features because all features contribute to the similarity
matching. Backtracking is also a challenging issue in nearest
kd-tree indexing algorithm.
(c) Right leaf node on left subtree
(d) Left leaf node on nearest kd-tree
Fig. 3 Illustration of Nearest kd-tree algorithm
To understand the drawbacks of kd-tree indexing algorithm,
consider the following example of searching the set of samples
shown in figure 3(a). The query candidate set is indicated in
red, the closest candidate set in yellow and the nearest
neighbors in blue. At first, the branch of the tree that is closest
to the query point (left subtree) is explored as shown in figure
3(b). At the next stage, the leaf nodes of the subtree are
traversed. (Fig. 3(c)). When a leaf is reached, the distance to
each point in the node is computed to check for a match. If it
is large, as in figure 3(d), then, backtracking is done and the
other branch at each node visited are tried. This backtracking
results in more search and retrieval time limiting the
performance of the kd-tree indexing algorithm.
One of the main reasons that nearest search on kd-tree
perform poorly is when there are more number of candidate
sets and the best candidate point can be either far from the
query point q or else close to the query point q . Hence, the
objective of optimising the kd-tree is to reduce the expected
number of visited leaves so as to improve the performance.
The main advantage of the optimisedkd-tree [15] is parameter
adjustability, that is, the direction and position of the split axis
as well as the maximal number of points.
B. Proposed Optimised kd-tree indexing algorithmfor fused
biometric database
This paper proposes an approach which tries to exploit the
Vol. 15, No. 8, Augus 2017
ISSN 1947-5500

optimisedkd-tree indexing technique combined with judicious
choice of index features for a multibiometric database. To
address the issues of the nearest kd-tree algorithm, extended
priority search is used where the priority search is
implemented among multiple trees to overcome the problem
of backtrack significantly. In the existing kd-tree algorithm,
searching may require that each data vector be examined. To
find the nearest neighbor of each point in a data set of N
vectors requires the comparison of N(N-1)/2 distances when
using a brute force method.
To optimize the kd-tree using extended priority search, the
priority level based on radius of hyper plane is fixed. Figure 4
shows the proposed optimisedkd-tree structure.
The innermost red circle indicates level 1 priority and
outermost green circle indicates level 2 priority. If the best
candidate falls on red hyper sphere then search is stopped else
it will move to outermost circle based on radius threshold
computation.
Fig. 4 Optimised kd-tree structure
The radius threshold for query candidate set q and neighbor
candidate set p (leaf node) is given by
(1)
Where, is outermost circle radius threshold, q is query
candidate set, p is end node (leaf) from nearest neighbors and
is inner most circle radius threshold.
C. Algorithmfor optimized kd tree
Fig. 5 Flowchart for optimized kd-tree
The optimized kd-tree indexing algorithm involves the
following steps:
1: Build a kd-tree for given templates from
multibiometric database.
2: For the given query q, perform nearest neighbor
search on kd-tree.
3: A nearest search first descends the tree and finds the
node that contains the query point as the first
candidate. Then it will find the nearest candidate p
based on the Euclidean distance limit which will be in
the outermost circle.
4: A point contained in this outermost circle is often not
the closest neighbor.
5: A priority search proceeds through other nodes of the
tree in order of their distance from the query.
6: The priority search is bounded by a hyper sphere
(innermost circle) of radius
7: The radius is adjusted when a better candidate is
found. When there are no more nodes within this
radius, the search terminates.
The flow chart of the optimized kd-tree algorithm is shown
in Fig 5.
III. PROPOSED OPTIMISED KD-TREE INDEXING ALGORITHM
WITH DIMENSION REDUCTION
To improve the efficiency of the proposed system model,
dimension reduction of the feature vectors is done. This
provides the advantages of lesser storage space and faster
retrieval from the large database. The dimension reduction of
the extracted features is achieved in this paper with the help of
popular Principal Component Analysis (PCA) algorithm.[18].
The main advantage of this reduction is to reduce the storage
space of template and also to ease the matching algorithm.
A multidimensional feature vector of each trait is first
normalized and then, it is projected to a lower dimensional
feature space. The reduced dimensional feature vectors are
fused at feature level and the fused feature vectors are used to
index the database by forming optimised Kd-tree. The
proposed method reduces the data retrieval time along with
possible error rates compared to optimised kd-tree without
dimension reduction using PCA.
A. Dimension reduction with PCA
The given two feature vectors FI and FF of d dimension that
belongs to Iris and Fingerprint respectively are reduced to s
dimension using the PCA. The features extracted from two
Vol. 15, No. 8, Augus 2017
ISSN 1947-5500

traits are of very high dimensions. Since all the features
generated are not equally important, proposed technique uses
PCA to reduce the dimensions which represents the compact
information. In the proposed technique, the two traits features
of finger print and iris are reduced from fdF and idI to fsF and isI
dimensions respectively.
The reduced s dimension features are fused to yield a
new feature vector VNto represent an individual. The vector
VNof dimensionality 2s, (2s≤ (fdF+idI)) can be generated by
augmenting vectors FF and FI of Fingerprint and iris
respectively. The fused feature values are then indexed using
the Optimised kd-tree.
B. Algorithmfor Optimised kd-tree with PCA
The enhanced Optimised kd-tree indexing algorithm including
the dimension reduction stage involves the following steps:
1: Normalize the feature vectors of each trait
2: Reduce the feature vectors using PCA.
3: Fuse the feature vectors at feature level fusion.
4: Form the Optimised kd-tree with the fused feature
vectors for indexing.
5: Invoke this indexing technique through multimodal
identification system.
In this technique, reduction makes use of Optimised kd-
tree structure for organizing the data in such a way that each
node can store one biometric template.
IV. SIMULATION RESULTS AND DISCUSSION
Simulation of the proposed model has been carried out
using 250 fingerprint images from the FVC database. The
fingerprint images were first subjected to preprocessing
techniques which include, selecting the region of interest,
histogramequalization and normalization. Finally the minutiae
features are extracted as per [16-17] and stored The extracted
features for given image is shown in figure 6.
(a)
(b)
Fig. 6 Minutiae Extraction (a) Skelton image (b) Minutiae points
The iris image is loaded from casia iris V4 database and
subjected to preprocessing to remove noise. Iris recognition
system comprises of four stages, iris segmentation, iris
normalization, and feature extraction and matching.
Iris segmentation separates an iris region from the entire
captured eye image by using circular Hough transform and
Canny edge detection. Iris normalization is performed to fix
the dimensions of segmented iris region to allow for accurate
comparisons by using Daugman’s Rubber Sheet Model.
(Fig.7)
Fig. 7 Iris Texture Feature
The features extracted individually from the two biometric
traits namely Fingerprint and Iris are fused using Feature level
fusion technique as shown in Figure 8(a).
The feature level fusion is advantageous over other fusion
techniques because of its better performance [18]. Figure 8(b)
shows the fused feature obtained by applying PCA for the
individual features.
Figure 8(a). Fusion of Fingerprint and Iris Features
Figure 8(b). Fusion of Fingerprint andIris features after PCA reductionby
54%
Next, the statistical feature of a fused image namely, mean
and standard deviation are computed and this is used as the
two dimensional key f(x,y) for improving the retrieval
performance for the multibiometric database.
Vol. 15, No. 8, Augus 2017
ISSN 1947-5500

A. Nearest kd tree indexing algorithm
For the given set of samples, the traditional kd-tree
structure is implemented which searches N closest point based
on input of query biometric image.
This query image has two dimensional key f(x,y). As
shown in figure 9, the blue line indicates x axis tree splitting
and red line indicates y axis tree splitting and this is mainly
based on depth of the tree (i.e level) in x or y in alteration
fashion. The pointswitin the circle indicate the candidate set
for the given query indicated by blue star.
NN search
Fig. 9. Nearest kd-tree search
B. Optimised kd-tree algorithm
Fig. 10 Optimised kd-tree search
For the given set of samples, the optimized kd-tree
structure is implemented which searches N closest point based
on distance between input of query candidate set and neighbor
candidate set (leaf node). Here once again, the blue line
indicates x axis tree splitting and red line indicates y axis tree
splitting. The innermost blue circle indicate level 1 priority
and outermost black circle indicate level 2 priority as shown in
figure 10. As explained in section 2.2 the priority is set based
on the radius threshold. For the given query q, a nearest search
first descends the tree and finds the node that contains the
query point as the first candidate.
Then it will find the nearest candidate p based on the
Euclidean distance limit which will be in the outermost circle.
A priority search proceeds through other nodes of the tree in
order of their distance from the query. If the best candidate
falls on innermost circle (blue) then search is stopped else it
will move to outermost circle based on threshold computation.
C. Performance Analysis
The performance of the proposed system is analysed by
testing the accuracy of matching between enrolment module
and verification module for a query template to identify an
individual for authorization. False Acceptance Rate and False
Rejection Rate are calculated based on genuine and imposter
authentication during verification. False Acceptance Rate
(FAR) is defined the number of times an imposter is
authenticated as genuine, as the criteria of reference threshold
is fulfilled. It is given by
False Rejection Rate (FRR) is defined as the number of
times a genuine person is rejected because the criteria of
reference threshold are not fulfilled. It is given by
The unibiometric (fingerprint and iris as individual
biometrics) and fused multibiometric indexing methods are
compared in terms of FAR and FRR. Typically, FAR and
FRR should be low which means imposter is not allowed to
authenticate and only genuine user is allowed access.
From the simulated results it is inferred that multibiometric
indexing has less FAR when compared to unibiometric and
FRR is comparatively reduced to 62.5% as shown in figure 11.
Fig. 11 Comparison of Unibiometric and Multibiometric Indexing
This indicates the advantage of using fused multibiometric
template in comparison to a single biometric system. The FAR
Vol. 15, No. 8, Augus 2017
ISSN 1947-5500

is also known as False Matching Ratio and FRR also known as
False Non-Matching Ratio. Thus, these two metrics help to
measure the accuracy of identification and security level of the
system.
D. Hit Rate and Penetration Rate
The nearest kd-tree and optimised kd-tree indexing
technique are compared in terms of penetration rate and hit
rate. A query image is said to be correctly identified, if one of
the retrieved identities corresponds to the correct identity. Hit
rate is defined as
Where C is the number of query images correctly identified
and P is the number of query images in the probe set.
The Penetration rate is the percentage of entries that have to
be retrieved by the indexing scheme for identification of probe
entries
Where N is the number of tests, Li is the number of images
in the candidate set of ith input image and M is the number of
identities in the database.
From the simulated results it is inferred that optimised kd-
tree indexing has more hit rate when compared to nearest kd-
tree and penetration rate is comparatively reduced as shown in
figure 12.
0 10 20 30 40 50 60 70 80 90 100
35
40
45
50
55
60
65
70
75
80
85
Performance of Indexed Database
Penetration rate (%)
Hitrate(%)
optimised kd tree technique
kd tree technique
Fig. 12 Hit rate versus Penetration rate
Table 1 shows the comparison between nearest kd-tree and
optimized kd-tree and dimension reduced features in
optimized kd-tree algorithms for the multibiometric database.
If retrieving process take penetration rate of 20%, 40%, 60%,
and 80%, the hit rate of the proposed technique is improved by
nearly 6% compared to the nearest kd-tree indexing technique
with less penetration rate and it also reduces search time
compared to nearest kd-tree indexing algorithm.
Thus the proposed optimisedkd-tree indexing technique
achieves 75% of reduction in False Acceptance Rate and 62.5
% of reduction in False Rejection Rate which outperforms
nearest kd-tree indexing technique in terms of accuracy of
identification and security. When the feature vectors of
reduced dimension after PCA is used to formthe optimized kd
tree, the hit rate performance is still better.
TABLE 1.
COMPARISON OF HIT RATE AND PENETRATION RATE
Penetration rate
(%)
20 40 60 80
Hit rate
(Nearest kd-tree
indexing )
57.5 65.5 70.5 74.0
Hit rate
(Optimisedkd-tree
indexing)
63.0 70.0 75.0 78.5
Hitrate (optimized
kd-tree indexing
with PCA)
76.4 80 85.9 89.6
Fig. 13 Comparison of performance with and without PCA
Figure 13 reveals that the hit rate of multibiometric
optimised kd-tree indexing with dimension reduction using
PCA increased at approximately 20% at different penetration
rates compared to the optimised kd-tree indexing without
dimension reduction using PCA. This shows the merit of the
proposed work.
V. CONCLUSIONS
In this paper, an efficient indexing and retrieving technique
based on optimisedkd-tree with dimension reduction using
PCA has been proposed for multibiometric identification.
Hit rate = (4)
Penetration rate = (5)
10 20 30 40 50 60 70 80 90 100
60
65
70
75
80
85
90
95
100
Penetration Rate(%)
HitRate(%)
Optimised kd tree indexing with PCA
Optimised kd tree indexing without PCA
Vol. 15, No. 8, Augus 2017
ISSN 1947-5500

Multibiometric databases require large storage due to huge
amount of templates (for example Aadhar database, ATM
database and other biometric databases). These types of
databases require efficient indexing to improve retrieval
efficiency. The experimental results obtained indicates that the
proposed indexing technique not only reduces the search space
and time but also gives good identification accuracy in terms
of hit rate and penetration rate when compared to
identification accuracies obtained by conventional database
search. The performance results is improved after use of PCA
as it reduces the memory space and reduces time required for
matching. Future research can also be done to analyze the
indexing performance where the database contains multiple
samples of a user from different types of sensors.
REFERENCES
[1] M.Liu and Pew-Thian yap. “Invariant representation of orientation
fields for finger print indexing,” Journal on Pattern recognition, vol.
45, pp. 2532-2542, January 2012.
[2] V.D.Kaushik, J.Umarani, A.K.Gupta, AmanK.Gupta, andP.Gupta. “An
efficient indexing scheme for face database using modified geometric
hashing,” Journalof Neurocomputing, vol.116, pp. 208-221, 2013.
[3] J.Umarani, AmanK.Gupta, and P.Gupta. “Boosted geometric hashing
based indexing technique for finger-knuckle-print,” Journal of
InformationSciences, vol. 275,pp.30-44, 2014.
[4] A.Gyaourova, and Arun Ross. “Index Codes for Multibiometric Pattern
Retrieval,” IEEE Transactions on Information Forensics and Security,
vol. 7, no. 2, April 2012.
[5] P.Mansukhani, S.Tulyakov, and V.Govindaraj. “A framework for
efficient fingerprint identification using a minutiae tree,” IEEE Systems
Journal, vol. 4, no. 2,pp.126-137, 2010.
[6] SomnathDeyansd and DebasisSamanta, “Iris data indexing method
using Gabor energy features,” IEEE Transactions on Information
Forensics and Security, vol. 7, no. 4,pp.1192-1203, 2012.
[7] AsimBaig, AlumedBouridane , Faith Kurugollu and Gang Qu.
“Fingerprint Iris Fusion Based identification system using a single
hamming distance Matcher,” International Journal of Bio-Science and
Bio-Technology, no. 1, pp. 47-57, 2009.
[8] Conti V, Milici G, Ribino P, Vitabile S and Sorbello F. “A frequency-
based approach for feature fusion in fingerprint and iris multimodal
biometric identification system,” IEEE Transactions on System, Man
and Cybernetics Part C: Applications and Reviews, vol. 40, no. 4, pp.
384-395,2010.
[9] Jain, A.K., Ross, A and Prabhakar, S. “An introduction to biometric
recognition,” IEEE Transactions on Circuits and Systems or Video
Technology , vol. 14, no. 1, pp. 4-20, 2004.
[10] J.Umarani, S.Prakash, Devdatt, and P.Gupta. “An indexing technique
for biometric database,” Proceedings of International Conference on
Wavelet Analysis and Pattern Recognition,vol.31,pp.758-763,2008.
[11] A.M.Briseno, A.G.Alonso, and J.H.Palanca. “Fingerprint indexing with
bad quality areas,” Journal on Expert system with applications, vol. 40,
pp. 1839-1846, 2013.
[12] J. Daugman. “The importance of being random: statistical principles of
iris recognition,”PatternRecognition, 36(2):279–291, February2003.
[13] J.Daugman. “High confidence visual recognition of persons by a test of
statistical independence,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 15, no. 11, pp1148-1160, 2011.
[14] HunnyMehrotra, BanshidharMajhi ,and Phalguni Gupta. “Robust iris
indexing scheme using geometric hashing of SIFT keypoints,” Journal
of Network and Computer Applications,vol.5, no. 4, 2010.
[15] ChanopSilpa-Anan, and Richard Hartley. “Optimisedkd-trees for fast
image descriptor matching,” IEEE Transaction on information
technology,vol.22,no.2, 2008.
[16] Gnana Kumar D, G.F.Sudha, and Revathi B. “An Efficient Space
Partitioning Tree Approach for Indexing and Retrieving Fingerprint
Databases,” Proceedings of ACM International Conference on
Advanced Research in Computer Science Engineering & Technology
ICARCSET’15, no. 50, pp. 259-263,March06-07,2015, India.
[17] Lukasz Wieclaw. “A minutiae-based matching algorithm,” Journal of
Medical Informatics and Technologies, vol. 13, pp. 65-71, 2009.
[18] Jayaraman U, Prakash S, Gupta P. “Indexing multimodal biometric
databases using kd-tree with feature level fusion,” Proceedings of the
4th International conference on Information systems security. Berlin,
Heidelberg: Springer; 2008.pp.221–234.
[19] Anitha, M. L., and K. A. RadhakrishnaRao. “An efficient palmprint
identification system based on an indexing approach,” International
Conference on Advances in Computing Communications and
Informatics (ICACCI), 2013.
AUTHORS PROFILE
Ms B.Revathi alias Ponmozhi receivedher B.Tech. in
ECE from Pondicherry University and M.E. degree in
Communication systems from the Anna University,
India, in 2000 and 2007, respectively. Currently, she is
an associate professor in the Department of Electronics
and Communication Engineering at Dhanalakshmi
Srnivasan Engineering College, Tamilnadu. She has more than 15 years of
experience in teaching.
She has published about 12 journal and conference papers. Her research
interest covers Biometric Security, Communication systems, and Pattern
Recognition. She is a life member of Indian Societyof Technical Education.
E-mail: brevathi@pec.edu
ORCID iD: 0000-0001-5039-3402
Dr.Gnanou Florence Sudha is currently working as Professor in the Dept. of
Electronics & Communication Engineering in
Pondicherry Engineering College affiliated to
Pondicherry University, Puducherry, India. She has
completed her B.Tech and M.Tech in Electronics and
Communication Engineering in 1992 and 1994. She
completed her PhD with specialization in Image
Processingin 2005.
She has more than two decade experience in teaching. She has several
journal and conference proceedings to her credit. She is member of the Indian
Society of Technical Education, Optical and Biomedical Society of India. Her
areas of research and interest are Image Processing, Biomedical signal
processing and Information Security. She has carried several research project
with All India Council for Technical Education and Dept. of Science and
Technology, India.
E-mail: gfsudha@pec.edu
Mr. D.Gnana Kumar is currently pursuing research in
the Dept. of Electronics & Communication Engineering
at National Institute of Technology, Tiruchirappalli,
Tamil Nadu, India. He has completed his BE in
Electronics and Communication Engineering in 2013 in
Anna university and M.Tech in wireless communication in 2015 in
Pondicherry University. His research interests are Database Management,
Image Processing, Wireless communication, Embedded Systems.etc. He has 6
Months experiencein teaching.
E-mail: gnanakumar421@gmail.com
Vol. 15, No. 8, Augus 2017
ISSN 1947-5500

Optimised Kd-Tree Approach with Dimension Reduction for Efficient Indexing and Retrieval from Multibiometric Database

Recommended

More Related Content

What's hot (20)

Similar to Optimised Kd-Tree Approach with Dimension Reduction for Efficient Indexing and Retrieval from Multibiometric Database (20)

Recently uploaded (20)

Optimised Kd-Tree Approach with Dimension Reduction for Efficient Indexing and Retrieval from Multibiometric Database