Documente Academic
Documente Profesional
Documente Cultură
4/2011 13
extraction of reliable minutiae. A fast All minutiae may not be genuine and
fingerprint enhancement algorithm [10] is significant amount of spurious minutiae are
used which can adaptively improve the detected during extraction process. Therefore
clarity of ridge and valley structures of the a methodology is used to remove all the
input fingerprint image based on the spurious minutiae detected in the minutiae
estimated local ridge orientation and extraction process. A minutia is considered to
frequency. be spurious if it lies in the fingerprint border
The enhanced binary image is thinned and or forms small clusters in the middle area of
minutiae are extracted. The ridge ending and the fingerprint image.
ridge bifurcation are the minutiae considered.
Select density = ρmin + β (ρmax - ρmin) (6) study for effective classification of
fingerprint dataset are random forest, J48,
The two data point’s xi and xj are considered NB tree, Bagging, adaboost and cost
to be in the same cluster if the following sensitive classifier.
condition is satisfied. Random forest is an ensemble decision trees
that offers good predictive performance.
ρi – ρj <= select density (7) Random forests construct a series of tree-
based learners. Each base learner receives
where ρi is the density of xi and ρj is the different training set which are drawn
density of xj. The un-chosen set represents independently with replacement from the
the data points which have never been original learning set. Each tree predicts a
selected for processing. The various steps of class which is considered as a vote and the
the clustering algorithm is as follows forest selects the class which as the most
1. The number of clusters and un-chosen votes.
set is given has input to the process. J48 tree is a reimplementation of C4.5 tree.
2. Calculate the distance between every NB Tree is a hybrid between decision trees
two data points and find Dmin and Dmax. and Naïve Bayes. It creates trees whose
3. Specify the density radius according to leaves are Naïve Bayes classifiers for the
Dmin and Dmax.. instances that reach the leaf. When
4. Calculate the density of each data point constructing the tree, cross-validation is used
and find ρmin and ρmax. to decide whether a node should be split
5. Specify the select density according to further or a Naive Bayes model should be
ρmin and ρmax. used instead.
6. Select a data point xi which has never Ensembles of classifiers are groups of
been chosen. classifiers in which the individual classifiers
7. Select any data point xj from xj‘s predictions are combined to classify new
density set and check for the same samples. Bagging and boosting is ensemble
cluster condition of classifiers used to improve the
8. Delete xi from the un-chosen set. performance of classification.
9. Select the next data point for clustering In bagging, many bootstrap samples are
from the un-chosen set using the drawn from the available data set, and some
following criteria prediction method like decision tree is
a. If intersection of un-chosen and applied to each bootstrap sample, and then
density set is not empty, select the next the prediction of individual classifiers are
data from the intersection set. combined by simple voting to obtain the
b. Otherwise, select any data from overall prediction. Here REP trees are used
un-chosen set. as base classifier. REP Tree builds a decision
tree using information gain/variance
5 Classifier Framework reduction and prunes it using reduced-error
The experience shows that no single machine pruning.
learning scheme is appropriate for all data Boosting, like bagging, is a committee-based
mining problems. In this section, different approach that can be used to improve the
learning models are compared and the most accuracy of classification. In adaboost,
appropriate algorithm is selected. The decision stump is used as base classifier. The
WEKA [11] is an open source data mining Decision Stump is a one-level binary
package which provides a collection of decision trees that can generate tree using
machine learning algorithms. During categorical or numeric class. It deals
training, all base classifiers are evaluated by effectively with missing values.
cross-validation on the fingerprint dataset. The cost-sensitive learning addresses the
The various classifiers used for comparative issue of classification in the presence of
Informatica Economică vol. 15, no. 4/2011 19
varying costs associated with different types cost-sensitive classifiers. For cost-sensitive
of misclassification. In particular, learning, one typically specifies only the
misclassification errors typically have non- costs for the false positives (C10) and false
uniform costs. These misclassification costs negatives (C01) and assigns a cost of zero to
are often determined by the class associated the true positives (C11) and true negatives
with an example, such that for two-class (C00).
problems, the cost of a false positive Conceptually, the cost of labeling an example
prediction is not equal to the cost of a false incorrectly should always be greater than the
negative prediction. cost of labeling it correctly.
Table 2 specifies the cost matrix used for
Table 3 illustrates the comparative results of classifiers and NB trees. The cost sensitive
various classifiers like bagging, Adaboost, classifier produces a better result with errors
J48, Random Forest, Cost sensitive minimized.
Image Processing Vol. 15, no. 3, pp. identification using cluster algorithm”, in
767–776, Mar. 2006. Proc. 16th ICPR, Vol. 2, pp. 398-401,
[2] X. Qiu, Y. Tang and Dan Meng, “ A new Aug. 2002
fuzzy clustering method based on [9] A. Senior and R. Bolle, “Improved
distance and density”, IEEE Conference fingerprint matching by distortion
on System, Man & Cybernetics, Oct. removal,” IEICE Trans. Information
2002,Vol 7, pp 6- 9. System, Special issue on Biometrics, Vol.
[3] X. P. Luo, J. Tian, and Y.Wu, “A minutia E84-D, no. 7, pp. 825–831, Jul. 2001.
matching algorithm in fingerprint [10] L. Hong, Y. Wan, and A. K. Jain,
verification,” in Proc. 15th ICPR, Sep. “Fingerprint image enhancement:
2000, Vol. 6, pp. 1039–1045 algorithms and performance evaluation,”
[4] N. K. Ratha and R. M. Bolle, “Effect of IEEE Transaction on Pattern Analysis
controlled acquisition on fingerprint and Machine Intelligence, Vol. 20, no. 8,
matching,” in Proc. 14th ICPR, 1998, pp. 777–789, Aug. 1998.
Vol. 2, pp. 1659–1661. [11] The WEKA package,
[5] A. K. Jain, L. Hong, and R. Bolle, “On- http://www.cs.waikato.ac.nz/ml/weka/
line fingerprint verification,” IEEE [12] A. K. Jain, S. Prabhakar, L. Hong, and
Transaction on Pattern Analysis and S. Pankanti, “Filterbank-based fingerprint
Machine Intelligence, Vol. 19, no. 4, pp. matching,” IEEE Trans. Image Process.,
302–313, Apr. 1997. vol. 9, no. 5, pp. 846–859 , May 2000.
[6] D. Lee, K. Choi, and J. Kim, “A robust [13] A. M. Bazen and S. H. Gerez,
fingerprint matching algorithm using “Fingerprint matching by thin-plate
local alignment,” in Proc. 16th ICPR, spline modeling of elastic deformations,”
Aug. 2002, Vol. 3, pp. 803–806 Pattern Recognition., vol. 36, no. 8, pp.
[7] X. Chen, J. Tian, X. Yang, and Y. Zhang, 1859–1867, 2003.
“An Algorithm for Distorted Fingerprint [14] M.S.Khalil1, D. Muhammad, M.K.
Matching Based on Local Triangle Khan and K. Alghathbar, "Singular
Feature Set,” IEEE Transaction on points detection using fingerprint
Information Forensics & security, Vol. 1, orientation field reliability", International
pp. 169-177, June 2006 Journal of Physical Sciences Vol.5, pp
[8] R. Qun, T. Jie, H. Yuliang and C. 352-357, April 2010
Jiangang, “Automatic fingerprint