=Paper=
{{Paper
|id=None
|storemode=property
|title=GEFeWS: A Hybrid Genetic-Based Feature Weighting and Selection Algorithm for Multi-Biometric Recognition
|pdfUrl=https://ceur-ws.org/Vol-710/paper11.pdf
|volume=Vol-710
|dblpUrl=https://dblp.org/rec/conf/maics/AlfordPDBKAAS11
}}
==GEFeWS: A Hybrid Genetic-Based Feature Weighting and Selection Algorithm for Multi-Biometric Recognition==
https://ceur-ws.org/Vol-710/paper11.pdf
GEFeWS: A Hybrid Genetic-Based Feature Weighting and
Selection Algorithm for Multi-Biometric Recognition
Aniesha Alford+, Khary Popplewell#, Gerry Dozier#, Kelvin Bryant#, John Kelly+,
Josh Adams#, Tamirat Abegaz^, and Joseph Shelton#
Center for Advanced Studies in Identity Sciences (CASIS@A&T)
+
Electrical and Computer Engineering Department,
#
Computer Science Department
^
Computational Science and Engineering Department
North Carolina A & T State University
1601 E Market St., Greensboro, NC 27411
aalford@ncat.edu, ktpopple@ncat.edu, gvdozier@ncat.edu, ksbryant@ncat.edu, jck@ncat.edu,
jcadams2@ncat.edu, tamirat@programmer.net, jashelt1@ncat.edu
Abstract Gentile, Ratha, and Connell 2009; Mumtazah and Ahmad
In this paper, we investigate the use of a hybrid genetic 2007). The goal of feature selection is to reduce the
feature weighting and selection (GEFeWS) algorithm for dimensionality of a data set by discarding features that are
multi-biometric recognition. Our results show that inconsistent, irrelevant, or redundant; thus keeping those
GEFeWS is able to achieve higher recognition accuracies features that are more discriminative and contribute the
than using genetic-based feature selection (GEFeS) alone, most to recognition accuracy. Feature weighting is a more
while using significantly fewer features to achieve general case of feature selection, with each feature being
approximately the same accuracies as using genetic-based assigned a weight based on its relevance (Yang and
feature weighting (GEFeW). Honavar 1998).
Genetic and Evolutionary Computation (GEC) has been
Introduction utilized by researchers to optimize feature selection and
weighting (Hussein et al. 2001; Yang and Honavar 1998;
A biometric system is a pattern recognition system that Yu and Liu 2003; Tahir et al. 2006; Raymer et al. 2000)
uses physiological and behavioral traits, characteristics that and has also been used by the biometrics community to
are unique for every individual, to perform recognition optimize the recognition accuracy (Dozier et al. 2009;
(Jain, Ross, and Prabhakar 2004). The value of a biometric Adams et al. 2010; Giot, El-Abed, and Rosenberger 2010).
system depends largely on its ability to accurately The goal of GEC is to find the optimal or near optimal
authenticate an individual. Thus, the recognition accuracy solution to a problem, and typically works as follows. A
is a major concern and is a key area of research for the population of candidate solutions is generated randomly
biometrics community (Deepika and Kandaswamy 2009). and assigned a fitness based on a user-defined function.
Researchers have shown that biometric systems that use Using this fitness, members of the population are chosen
only one biometric modality can produce highly accurate and reproduce. The resulting offspring are then evaluated
results (Adams et al. 2010; Dozier et al. 2009; Miller et al. and typically replace candidate solutions within the
2010; Ross 2007). However, when these systems are population that have a lower fitness. This evolutionary
applied to real-world applications, their performance can process is continued until the population converges, a user-
be affected by numerous factors such as noisy sensor data specified number of evaluations have completed, or no
due to dust or lighting conditions and spoofing. Multi- solution can be found.
biometric systems that fuse multiple biometric modalities In this paper, we use a hybrid GEC-based feature
have been shown to be more robust, able to counter many weighting and selection (GEFeWS) technique for multi-
of the aforementioned limitations, and are also capable of biometric recognition. Our goal is to reduce the number of
achieving higher recognition accuracies (Jain, features necessary for biometric recognition and increase
Nandakumar, Ross 2005; Ross 2007; Eshwarappa and the recognition accuracy. The performance of GEFeWS is
Latte 2010). compared with the performances of genetic-based feature
Feature selection and weighting have also been proven selection (GEFeS) and weighting (GEFeW) techniques
as successful methods of improving the accuracy rates of individually. The modalities tested were face and
biometric systems (Adams et al. 2010; Dozier et al. 2009; periocular biometrics. The facial features were extracted
using the Eigenface method (Turk and Pentland 1991; Lata
et al. 2009), and the periocular features were extracted
using Local Binary Patterns (LBP) (Adams et al. 2010; Binary Patterns (LBP) is used to extract features from the
Miller et al. 2010). periocular region (Adams et al. 2010; Miller et al. 2010).
This research is inspired in part by the proposal of a Eigenface is based on the concept of Principal
hierarchical two-stage system, presented by Gentile et al. Component Analysis (PCA) and has been proven
for iris recognition (2009). This system used a reduced successful for facial recognition (Turk and Pentland 1991;
feature set size in an effort to reduce the total number of Lata et al. 2009). PCA is a method used to reduce the
feature checks required for an iris-based biometric dimensionality of a dataset while retaining most of the
recognition system. For a conventional biometric variation found among the data (Jolliffe 2005). For the
recognition system, a probe, p, is compared to every Eigenface method, PCA is used to find the principal
individual within a biometric database. The number of components, or eigenfaces, of the distribution of the face
feature checks performed by a conventional biometric images within the entire image space, which is called the
system, fc, is: face space.
LBP is a method used for texture analysis that has been
fc = nm used in many biometric applications, including the
extraction and analysis of periocular features for
where n is the number of individuals in the database and m identification (Adams et al. 2010; Miller et al. 2010). LBP
is the number of features used to represent an individual. descriptors of each periocular region are formed by first
A hierarchical biometric system reduces the number of segmenting the image into a grid of 24 evenly sized
feature checks performed by first using the reduced length patches. Every internal pixel within the patch is used as a
biometric template to select a subset of the r closest center pixel. The intensity change of the pixels around the
matches to the probe p. The subset is then compared to p center pixel is measured by subtracting the intensity value
using all of the m features. The number of feature checks of the center pixel from each of the P neighboring pixels.
performed by a hierarchical system, fh, is the summation of For our experiments, the neighborhood size, P, was 8. If
the calculations of the two stages, represented by: the resulting value is greater than or equal to 0, a 1 would
be concatenated to the binary string representing the
fh = nk + rm texture, otherwise a 0. The texture is then encoded into a
histogram where each bin represents the number of times a
where, once again, n represents the number of individuals particular binary string appears in a patch. For
in the database, k is the number of features in the reduced optimization purposes, only uniform patterns are
feature set, r is the subset of the closest r-individuals to the considered. These are binary string patterns with at most
probe, p, and m is the number of features used to represent two bitwise changes when the pattern is traversed
an individual. The savings gained by using the hierarchical circularly. Therefore, our histogram consisted of 59 bins
biometric system, fs, instead of the conventional biometric (instead of 2P=256 bins), 58 for the possible uniform
system is: patterns and 1 for the non-uniform patterns.
f nk + rm k r GEFeS, GEFeW, and GEFeWS
fs = h = = +
fc nm m n The genetic and evolutionary techniques used within this
paper are based on the eXploratory Toolset for the
Optimization of Launch and Space Systems (X-TOOLSS)
The remainder of this paper is as follows. In the
(Tinker, Dozier, and Garrett 2010), and are an instance of
following section, a brief overview of the feature extractors
the X-TOOLSS Steady-State Genetic Algorithm (SSGA).
used for our experiments is given. GEFeS, GEFeW, and
For GEFeS, a SSGA is used to evolve a feature mask
GEFeWS are then described, followed by a description of
that selects the most salient biometric features. For each
our experiments, the presentation of our results, and
real-valued candidate solution that is generated by the
finally, our conclusions and future work.
SSGA, a masking threshold of 0.5 is used to determine if
the feature is used. If the values of the features within the
Feature Extraction mask are less than the masking threshold, the feature is
turned off by setting the mask value to 0. Otherwise, the
Feature extraction is one of the essential tasks performed
feature is turned on by setting the mask value to 1,
by a biometric system. After a biometric sample is
resulting in a binary coded feature mask.
acquired from an individual, feature extraction is
For GEFeW, a SSGA is used to evolve a real-valued
performed to extract a set of features, termed a feature
feature mask composed of values between 0.0 and 1.0.
template, which is used to represent the individual and is
The resulting feature mask value is multiplied by each
used in the comparisons to determine recognition (Jain,
feature value to provide the weighted feature.
Ross, and Prabhakar 2004).
GEFeWS is a hybrid of GEFeW and GEFeS. Like
In this paper, we use two feature extraction schemes.
GEFeW, a SSGA is used to evolve the weight of the
The Eigenface method is used to extract features from the
features. However, if the weight is less than the masking
face (Turk and Pentland 1991; Lata et al. 2009). Local
threshold of 0.5, then the feature is not included, basically Table I shows the performance comparison of GEFeS,
being turned off as done by GEFeS. Otherwise, the feature GEFeW, and GEFeWS. The results using the feature
is weighted as done by GEFeW. extractors without the GECs were also included to serve as
Associated with each candidate feature mask, i, there a baseline for the experiments. When the face and
were two weights, wip and wif, which are weights for the periocular biometrics were fused, they both were weighted
periocular and face feature submasks to allow for score- evenly.
level fusion. The weights ranged from [0..1] and were co- For the Face-Only experiment, GEFeW performed the
evolved with the rest of the feature mask. best in terms of accuracy, having an average accuracy of
87.59%. Based on the results of the ANOVA and t-test,
Experiment GEFeWS was in the second equivalence class in terms of
average accuracy, but there was only a 1.21% difference in
To test our algorithms, we used a subset of 315 images the average accuracy for the two algorithms. In terms of
taken from the first 105 subjects of the Face Recognition the percentage of features used, GEFeWS was in the first
Grand Challenge (FRGC) dataset (Phillips et al. 2005). equivalence class, along with GEFeS. GEFeWS was able
These images were used to form a probe set of 105 images to obtain an average accuracy of 86.38%, while using only
(one of each subject) and a gallery set of 210 images (two 51.71% of the features.
of each subject). For each of the images in the probe and For the Periocular-Only experiment, GEFeWS
gallery set, the Eigenface method was used to extract 210 performed the best in terms of accuracy and the percentage
face features, and the LBP method was used to extract of features used, having an average accuracy of 96.15%
2832 periocular features (1416 features for each eye). while using only 45.39% of the features. These results
Three biometric modalities were tested: face, periocular, were confirmed using an ANOVA and t-test. GEFeW was
and face plus periocular. For each of the three biometric in the second equivalence class in terms of average
modalities, GEFeS, GEFeW, and GEFeWS were used. The accuracy. In terms of the percentage of features used,
biometric modalities were also tested using all of the GEFeS and GEFeW were in the second and third
originally extracted features without the use of GECs. This equivalence classes respectively.
served as a control/baseline for our experiments. For the Face + Periocular experiment, GEFeW
performed the best in terms of accuracy, while GEFeWS
Results was in the second equivalence class. However, in terms of
the percentage of features used, GEFeWS was in the first
For our experiments, the SSGA had a population size of 20 equivalence class, using only 46.24% of the features to
and a Gaussian mutation range of 0.2. The algorithm was achieve an average accuracy of 98.48% (only a 0.5%
run 30 times, and a maximum of 1000 evaluations were difference when compared to GEFeW). GEFeS and
performed on each run. GEFeW were in the second and third equivalence classes
In Table I, the average performance of the three respectively.
experiments is shown. The first column represents the The Face + Periocular experiment performed the best in
tested biometric modalities. The second column represents terms of accuracy for all the algorithms used, followed by
the type of algorithm that was used. The third column the Periocular-Only experiment and the Face-Only
represents the average percentage of features used, and the experiment.
last column represents the average accuracy of the 30 runs.
Modalities Average % of Average
Algorithms Used
Tested Features Used Accuracy
Eigenface 100.00% 64.76%
Face Eigenface + GEFeS 51.03% 77.87%
Only Eigenface + GEFeW 87.71% 87.59%
Eigenface + GEFeWS 51.71% 86.38%
LBP 100.00% 94.29%
Periocular LBP + GEFeS 48.03% 95.14%
Only LBP + GEFeW 86.22% 95.46%
LBP + GEFeWS 45.39% 96.15%
Eigenface + LBP [evenly fused] 100.00% 90.77%
Face + Eigenface + LBP + GEFeS 48.18% 97.40%
Periocular Eigenface + LBP + GEFeW 87.59% 98.98%
Eigenface + LBP + GEFeWS 46.24% 98.48%
Table 1. Comparison of the performances of GEFeS, GEFeW, and GEFeWS.
For the percentage of features used, GEFeWS used the Gentile, J.E.; Ratha, N.; and Connell, J. 2009. SLIC: Short-
least amount of features for the Periocular-Only and Face Length Iris Codes. In Proceedings of the IEEE 3rd International
+ Periocular experiments, and there was no statistical Conference on Biometrics: Theory, Applications, and System,
significance between GEFeWS and GEFeS for the Face- (BTAS).
Only experiment. GEFeW used the highest percentage of Gentile, J.E.; Ratha, N.; and Connell, J. 2009. An Efficient, Two-
features for all three experiments. stage Iris Recognition System. In Proceedings of the IEEE 3rd
International Conference on Biometrics: Theory, Applications,
and System, (BTAS).
Conclusion Giot, R.; El-Abed, M; and Rosenberger, C. 2010. Fast Learning
Our results show that the hybrid GEC, GEFeWS, is able to for Multibiometrics Systems Using Genetic Algorithms. In
Proceedings of the IEEE International Conference on High
achieve higher recognition accuracies than GEFeS, while
Performance Computing and Simulation (HPCS).
using about the same amount of features. GEFeWS is also
able to use a significantly lesser amount of features than Hussein, F.; Kharma, N.; and Ward, R. 2001. Genetic
Algorithms for Feature Selection and Weighting, a Review and
GEFeS while achieving approximately the same average
Study. In Proceedings of the Sixth International Conference on
recognition accuracy. Overall, the Face + Periocular Document Analysis and Recognition.
performed better in terms of accuracy when compared to
Jain, A.K.; Duin, R.P.W.; and Jianchang M. 2000. Statistical
the Face-Only and Periocular-Only experiments. Our
Pattern Recognition: A Review. In IEEE Transactions on Pattern
future work will include investigating additional multi- Analysis and Machine Intelligence, Volume (22), Issue (1): pp.
biometric fusion techniques as well as additional GECs in 4–37.
an effort to further improve the performance of multi-
Jain, A.K.; Ross, A.; and Prabhakar, S. 2004. An Introduction to
biometric recognition. In addition, we will investigate Biometric Recognition. In IEEE Transactions on Circuits And
applying these algorithms to a larger dataset to see how Systems for Video Technology, Volume (14), Issue (1): pp: 4-20.
well they generalize.
Jain, A.; Nandakumar, K.; and Ross, A. 2005. Score
Normalization in Multimodal Biometric Systems. Pattern
Acknowledgments Recognition, Volume (38), Issue (12): pp. 2270–2285.
Jolliffe, I. 2005. Principal Component Analysis. Encyclopedia of
This research was funded by the Office of the Director of Statistics in Behavioral Science.
National Intelligence (ODNI), Center for Academic
Kohavi, R.; Langley, P.; and Yun, Y. 1997. The Utility of
Excellence (CAE) for the multi-university Center for
Feature Weighting in Nearest-Neighbor Algorithms. In
Advanced Studies in Identity Sciences (CASIS) and by the Proceedings of the 9th European Conference on Machine
National Science Foundation (NSF) Science & Learning (ECML).
Technology Center: Bio/computational Evolution in
Lata, Y.V.; Tungathurthi, C.; Rao, R., Govardhan, A.; and Reddy,
Action CONsortium (BEACON). The authors would like L.P. 2009. Facial Recognition using Eigenfaces by PCA. In
to thank the ODNI and the NSF for their support of this International Journal of Recent Trends in Engineering, Volume
research. (1), Issue (1): pp. 587-590.
Miller, P.E.; Rawls, A.; Pundlik, S.; and Woodard, D. 2010.
References Personal Identification Using Periocular Skin Texture. In
Proceedings of the 2010 ACM Symposium on Applied Computing.
Adams, J.; Woodard, D.L.; Dozier, G.; Miller, P.; Bryant, K.;
Mumtazah, S. and Ahmad, S. 2007. A Hybrid Feature Weighting
and Glenn, G. 2010. Genetic-Based Type II Feature Extraction
and Feature Selection Approach in an Attempt to Increase
for Periocular Biometric Recognition: Less is More. In
Signature Biometrics Accuracy. In Proceedings of the
Proceedings of the 20th International Conference on Pattern
International Conference on Electrical Engineering and
Recognition.
Informatics.
Deepika, C.L. and Kandaswamy, A. 2009. An Algorithm for
Phillips, P.J.; Flynn, P.J.; Scruggs, T.; Bowyer, K.W.; Chang, J.;
Improved Accuracy in Unimodal Biometric Systems through
Hoff, K.; Marques, J.; Min, J.; and Worek, W. 2005. Overview
Fusion of Multiple Feature Sets, In ICGST International Journal
of the Face Recognition Grand Challenge. In Proceedings of
on Graphics, Vision and Image Processing, (GVIP), Volume (9),
IEEE Conference on Computer Vision and Pattern Recognition.
Issue (III): pp. 33-40.
Raymer, M.L.; Punch, W.F.; Goodman, E.D.; Kuhn, L.A.;
Dozier, G.; Frederiksen, K.; Meeks, R.; Savvides, M.; Bryant, K.;
Jain, A.K. 2000. Dimensionality Reduction Using Genetic
Hopes, D.; and Munemoto, T. 2009. Minimizing the Number of
Algorithm. In IEEE Transactions on Evolutionary Computation,
Bits Needed for Iris Recognition via Bit Inconsistency and GRIT.
Volume (4), Issue (2): pp: 164-171.
In Proceedings of the IEEE Workshop on Computational
Intelligence in Biometrics Theory, Algorithms, and Applications, Ross, A. 2007. An Introduction to Multibiometrics. In
(CIB). Proceedings of the 15th European Signal Processing Conference
(EUSIPCO).
Eshwarappa M.N. and Latte, M.V. 2010. Bimodal Biometric
Person Authentication System Using Speech and Signature Tahir, M.A., et al., 2006. Simultaneous Feature Selection and
Features. In Proceedings of the International Journal of Feature Weighting using Hybrid Tabu Search/K-Nearest
Biometrics and Bioinformatics, (IJBB), Volume (4), Issue (4). Neighbor Classifier. Pattern Recognition Letters. Volume (28),
Issue (4): pp. 438-446.
Tinker, M.L.; Dozier, G.; and Garrett, A. 2010. The eXploratory Yang, J. and Honavar, V. 1998. Feature Subset Selection Using a
Toolset for the Optimization Of Launch and Space Systems X- Genetic Algorithm, In Proceedings of the IEEE Intelligent
TOOLSS). http://xtoolss.msfc.nasa.gov/. Systems and their Applications. Volume (13), Issue: (2): pp. 44-
Turk, M. and Pentland, A. 1991. Eigenfaces for Recognition. 49.
Journal of Cognitive Neuroscience. Volume (3), Issue (1): pp. Yu, L. and Liu, H. 2003. Feature Selection for High-Dimensional
76-81. Data: A Fast Correlation-Based Filter Solution, In Proceedings of
the Twentieth International Conference on Machine Learning.