close

Вход

Забыли?

вход по аккаунту

?

978-981-10-6427-2 42

код для вставкиСкачать
Brain Signal Based Biometric Identification
Using One-Dimensional Local Gradient Pattern
and Artificial Neural Network
Abeg Kumar Jaiswal(B) and Haider Banka
Department of Computer Science and Engineering,
Indian Institute of Technology (ISM) Dhanbad,
Dhanbad 826004, Jharkhand, India
abegiitdhanbad@gmail.com, hbanka20002@yahoo.com
Abstract. Biometric identification or recognition refers to the process
of identifying an unknown individual based on the physiological or behavioral characteristics. While fingerprint, palm, face belongs to physiological characteristics, traits like voice, gait falls in the category of behavioral
characteristics. Recently, there has been an increase in interest in developing neural signal based biometric identification system as the brain
signals have certain unique features related to an individual and they
are difficult to mimic as well. Electroencephalogram (EEG) captures
the brain electrical activity and used in different applications including
health care and human-computer interaction. In this paper, a new approach with the combination of one-dimensional Local Gradient Pattern
(1D-LGP) and Artificial Neural Network (ANN) has been introduced
for building EEG signal based biometric identification system. The proposed framework consists of two steps. In the first step, the 1D-LGP
code is computed for each signal point in the EEG signal and the histogram is formed. The histogram represents the extracted feature vector
of the corresponding EEG signal which is then fed to the ANN classifier
to perform the classification. The experiment has been carried-out with
the benchmark dataset having 20 subjects. The system performance has
been evaluated using the mean accuracy obtained after 20 runs of 10fold cross validation. The experimental results show that the proposed
technique achieved a high accuracy.
Keywords: Electroencephalogram (EEG) signals · One dimensional
Local Gradient Pattern (1D-LGP) · Artificial Neural Network (ANN)
1
Introduction
The brain consists of approximately 90–100 billion neurons. Electroencephalogram (EEG) captures the electrical activities of these neurons by placing electrodes on the scalp. This capability of EEG signal makes it an important tool
for application in different domains of brain research. EEG signals are used in
c Springer Nature Singapore Pte Ltd. 2017
J.K. Mandal et al. (Eds.): CICBA 2017, Part I, CCIS 775, pp. 525–536, 2017.
DOI: 10.1007/978-981-10-6427-2 42
526
A.K. Jaiswal and H. Banka
different applications, such as diagnosis of epilepsy and building biometric identification system. Recently, there has been an increase in interest in building EEG
based biometric identification system. In biometric identification system, the
input sample of an unknown introducer is matched with the samples of several
individuals of the dataset. The best match reveals the identity. The biometric
identification is considered as a multi-class classification problem.
In the last few years, a number of frameworks have been proposed by
researchers for EEG signal classification. The two basic steps involved in these
frameworks are feature extraction and classification. Feature extraction reduced
the dimension of the input patterns by keeping the most important attribute
and constitute the feature vectors which are then given as input to a classifier
to carry out the classification. This process is usually followed while building an
automated signal classification system.
In recent years, several biometrics including face, fingerprint, and eye are in use
for identifying individuals. EEG has also been reported to have certain biometric
characteristics and a number of techniques have been proposed for development
of EEG based biometric system [1]. Some of the popular techniques includes maximum a posteriori model [2], multiple signal classification (MUSIC) algorithm [3],
channel energy based neural network [4], face specific visualization with biserial
correlation coefficient [5], Hilbert-Huang transform [6], time-frequency analysis
with power spectrum density [7], phase synchronization based eigenvector centrality technique [8], multi-linear PCA (MPCA) [9], Binary Flower Pollination
Algorithm [10], and wavelet transform based neural network [11].
Local binary pattern (LBP) has gained popularity in different pattern recognition applications including face recognition [12]. Later, One dimensional LBP
(1D-LBP) was proposed for signal processing [13] and recently, successfully
applied for epileptic EEG signal classification [14,15]. However, it has certain
limitation.
In this study, a new framework by combining One dimensional Local Gradient Pattern (1D-LGP) and Artificial Neural Network (ANN) is introduced
for EEG based biometric identification. Both, LBP and LGP have been used
for face recognition [12,17,18]. Both these techniques are well known for their
computational simplicity. Each action or abnormality recorded in EEG signal
posses some unique patterns. 1D-LBP focus on the local pattern structure and
can detect these hidden patterns. Recently, 1D-LBP was successfully applied for
feature extraction in non-stationary signals like EEG signals [14,15]. However,
1D-LBP is sensitive to local variation. Whereas, the 1D-LGP technique is insensitive to both local and global variations [16]. 1D-LGP consists of two steps.
In the first step the 1D-LGP code is computed for each signal point and then
these codes are used to form the histogram in the second step. The histogram
represents extracted feature vector of the corresponding EEG signals. This histogram contains the structural distribution of patterns across the EEG signal
and fed to a classifier to carry-out the classification. In this study, ANN classifier
has been used and the classification accuracy is computed through 10-fold cross
validation.
Brain Signal Based Biometric Identification
527
The rest of the paper is organized as follows. Section 2 describe the methodology and different datasets used for the research. The Experimental outcomes are
presented in Sect. 3. Finally, Sect. 4 concludes the article with future direction.
2
Methodology and Materials
This section describes the LBP, LGP, 1D-LGP techniques, ANN classifier and
the dataset used in this research.
2.1
Local Binary Pattern (LBP) and Local Gradient Pattern (LGP)
LBP and LGP are feature extraction techniques used in face recognition [12,18].
In LBP, the face image pixels are represented by the relationship between the center pixel and its surrounding pixels directly. On the other hand, in LGP technique,
pixels are represented by the their relationship between the gradient values. In case
of image, the gradient value of a pixel is computed from the gradient values of 8
neighbor pixels surrounding it. The gradient value is given by the absolute difference between the intensity value of the given pixel and its neighboring pixel. Once
all the eight gradient values are obtained, the average gradient value of the surrounding neighborhood pixels is computed, it is then considered as the threshold
in LGP for that pixel. Each of the 8 gradient values are compared with the average gradient value. If any gradient value is greater that the threshold than that
corresponding neighboring pixel value is assigned 1, or else it is assigned 0. The
final LGP code of any given pixel is obtained by the concatenation of the binary
code values. One of the examples of the LBP and LGP code computation is shown
in Fig. 1.
Fig. 1. LBP and LGP code for the center pixel.
LGP can also be considered to have different size of the neighborhood, where
n represents the radius of the neighborhoods and m represent number the pixel
points under consideration.
528
2.2
A.K. Jaiswal and H. Banka
Proposed Model
Figure 2 depicts the structure of the EEG based biometric identification system.
EEG Signal
Classifier
1D -LGP
Identification
ANN
1D-LGP code
Histogram
Subject
Subjects
Fig. 2. EEG based biometric identification.
2.3
One Dimensional Local Gradient Pattern (1D-LGP)
The various steps involved in the 1D-LGP technique are as follows:
1. Consider m neighboring points.
2. Mark m/2 points in the left and right of the center point Sc .
3. Once the above two steps are over, the gradient value gi is computed.
gi = |Pi − Sc |, for i = 0, ..., m − 1.
4. In the fourth step, the average (gavg ) of all the gradient value is calculated.
m
gavg =
1 gi
m i=1
(1)
5. The gradient codes (gci ) are obtained by subtracting the average (gavg ) from
the individual gradient value (gi ). gci = gi − gavg , for i = 0, ..., m − 1.
6. Finally, the 1D-LGP code for the center point Sc is given by:
m−1
Sc 1D−LGP =
s(gci )2i
(2)
i=0
where,
1,
s(x) =
0,
if x ≥ 0
otherwise
The various steps of the 1D-LGP feature extraction technique are explained
with the small segment of an EEG signal as shown in Fig. 3.
In case of 1D-LBP [13,14] the transformation code is computed as:
Sc
1D−LBP
=
m−1
i=0
s(Pi − Sc )2i
(3)
Brain Signal Based Biometric Identification
529
Fig. 3. Computation of 1D-LGP code for EEG signal point Sc.
Once the transformation code computation is over for the all signal points,
the histogram is formed using these codes. This histogram represents the feature
vector of the EEG signal and used for the classification. The both the techniques
(1D-LBP and 1D-LGP), the length (l) of the histogram based feature vector
considering m neighboring points is given as:
l = 2m
(4)
The transformation code acquires a value between 0 to 2m−1 (inclusive).
2.4
Artificial Neural Network (ANN)
Artificial Neural Network (ANN) is a machine learning classifier and widely
used in different pattern recognition applications. There are different ANNs. In
this study, we have used the multilayer preceptron (MLP) neural network. The
MLP consists of three layers, an input layer, hidden layer and the output layer.
Here, each layer is fully connected to the next layer unidirectionally. Each layer
consists of neurons. The neurons receive data from its previous layer, process the
data and passes the processed data to the neurons in the successive layer. The
processing of data is performed by weight and activation function. Usually, the
activation function is nonlinear in nature. MLP uses back propagation technique
for training the network. A detailed description about MLP can be found in [19].
530
A.K. Jaiswal and H. Banka
2.5
1D-LBP and 1D-LGP in Case of Local and Global Variations
Tolerance to noise is one of the most important properties of a feature extraction
technique. Usually noise is present in most of the acquired signals. A noise my
cause a local or a global variation. 1D-LGP is insensitive to local and global
variations. The behavior of 1D-LGP and 1D-LBP in case of local and global
variations is shown in Fig. 4.
EEG segment:
(a)
240
243
114
115
116
242
246
113
239
1
0
1
1d-LGP
1d-LGP
1D-LGP
1
1
0
0
1
Global Variation
162
165
36
37
38
164
Local Variation
168
35
161
240
243
124
125
1
0
1
1
1
0
0
246
113
239
0
1
116
1d-LGP
1
0
0
1
246
113
239
1
0
1
113
239
1d-LGP
1D-LGP
1D-LGP
1
242
1
(b)
240
243
114
1
1
0
115
116
242
1d-LBP
1D-LBP
0
1
1
Global Variation
162
165
36
37
38
164
Local Variation
168
35
161
240
243
124
125
1D-LBP
1d-LBP
1
1
0
0
1
1
116
242
246
1D-LBP
1d-LBP
0
1
1
1
1
1
1
1
0
1
Fig. 4. 1D-LGP and 1D-LBP in case of local and global variations.
In the absence of noise, the patterns with similar structural properties are
represented by same 1D-LGP code. This makes 1D-LGP a better feature representation technique than other feature representation technique like 1D-LBP. It
can be seen from the Fig. 4 that the patterns of EEG signals are represented by
the same 1D-LGP code in case of both local and global amplitude variations,
whereas 1D-LBP is sensitive to local variation.
2.6
Dataset
This dataset1 consists of 20 individuals (10 alcoholic and 10 control subjects)
and is provided by Neurodynamics Laboratory, State University of New York
Health Center Brooklyn, New York. There are 40 EEG signals of each individual.
So the dataset contains a total of 800 EEG signals of 20 individuals. Here, the
set of signals of an individual corresponds to a particular class. The EEG signals
were recorded from 64 electrodes system and the sampling frequency was 256 Hz
1
EEG Dataset https://archive.ics.uci.edu/ml/datasets/EEG+Database.
Brain Signal Based Biometric Identification
531
per second. Each subject was shown some pictures selected from the Snodgrass
and Vanderwart picture set. The task of each individual was to recognize these
pictures while displayed on the CRT screen. To perform a fair experiment the
distance between the screen and individual was set to one meter. Each picture
was shown for 3 s. After a picture was shown, 1 s EEG signal was recorded. The
time interval between two consecutive pictures shown was set for approximately
5 s. One of the recorded signal is shown in Fig. 5.
50
Electrode 1
0
−50
0
50
50
100
150
200
250
50
100
150
200
250
50
100
150
200
250
50
100
150
200
250
Electrode 2
0
Amplitude (µV)
−50
50
0
Electrode 3
0
−50
0
50
Electrode 64
0
−50
0
Signal Points
Fig. 5. Recoded EEG signal.
So the number of features in each sample of this dataset is 64 * 256. The
dataset is described in Table 1.
Table 1. Dataset
Number of Number of samples Number of features Total number of
subjects
per subject
in each sample
samples
20
40
64 * 256
800
532
A.K. Jaiswal and H. Banka
The EEG recordings of different individuals are shown in Fig. 6.
20
Subject 1
0
0
−20
−20
50
Amplitude (µV)
Subject 2
20
20
100
150
200
Subject 3
20
0
20
150
200
Subject 4
0
−20
−40
100
−20
50
100
150
200
Subject 5
50
20
0
100
150
200
Subject 6
0
−20
−20
−40
100
150
200
50
100
150
200
Electrode 1 recording
Fig. 6. EEG recoding of electrode 1 of different subjects.
3
Experimental Results and Discussion
In this section, the experimental results have been shown and the analysis of
results has been carried out.
3.1
Results
In the 1D-LGP technique, the first step is to compute the 1D-LGP code for
each signal point. Once the code computation is over for the entire signal the
histogram is formed using the codes. The histogram formed the feature vector
of the corresponding EEG signal. These feature vectors are given as input to the
ANN classifier perform the classification. For visual representation, the small
segments of histogram based feature vector for different subjects are shown in
Fig. 7. Signals of the same individual will have similar histograms. On the other
hand, the histograms of different individuals will be different.
In this research, multilayer perceptron artificial neural network has been
used to carry out classification of EEG signals using MATLAB R2013a. The
maximum iteration for convergence was set to 1000 whereas the mean square
error was set to .0001. The scaled conjugate gradient method (trainscg) was
Brain Signal Based Biometric Identification
533
250
Count
200
150
100
50
0
121
122
123
1 124
D−
LG
125
P C 126
ode
127
128
129
130
1
2
4
3
5
ts
bjec
Su
Fig. 7. 1D-LGP histogram segments of 5 different subjects.
used along with the hyperbolic tangent sigmoid (tansig) transfer function. The
number of neurons in hidden layer was set to 55. The classification accuracy is
computed with the mean obtained from 20 iterations of 10-fold cross validation.
As mention before, in a biometric identification system the task is to find
the identity of an introducer by performing one-to-many comparison against the
biometric database. The dataset contains the recording of 20 subjects. In order
to extract the features from each EEG signals using 1D-LGP, the dimension
of each raw EEG signal was transformed by concatenating each channel data
row wise (64 * 256 = 1 * 16384). We apply the 1D-LGP technique (with m = 8)
on each transformed EEG signal to form the feature vector. Once the feature
extraction was over the extracted feature vectors were classified using ANN classifier. Table 2 shows the classification accuracy (Acc) and the standard deviation
(std) achieved by 1D-LBP and 1D-LGP with ANN classifier.
Table 2. Classification accuracy (%) of 1D-LBP and 1D-LGP with ANN classifier.
Number of classes Classifier Feature extraction technique
1D-LBP (Acc (%) ± Std) 1D-LGP (Acc (%) ± Std)
20
ANN
62.70 ± 2.65
88.90 ± 1.30
It can be seen in Table 2 that the classification accuracy achieved by the
proposed 1D-LGP technique is high as compared to 1D-LBP.
One of the major limitations in the EEG based biometric research is the
unavailability of a large public dataset. A number methods have been proposed
in the literature for building the EEG based biometric identification system using
different datasets with different number subjects. The experimental results of
various methods in literature and proposed techniques are shown in Table 3.
534
A.K. Jaiswal and H. Banka
Table 3. Some techniques proposed in literature for biometric identification system.
Authors Year Method
Subjects Accuracy
[20]
2010 WT + ANN
10
81
[21]
2010 AR + KNN
10
96.7
[5]
2013 PSD + ANN
50
95
[23]
2014 AR + SVM
20
97.25
[26]
2014 ET + LVQNN
6
90.97
[22]
2015 CNN
[25]
2016 WT + signal energy + NN
[24]
2016 WT+ ETB + LDA
10
88
5
95.00
30
83.00
20
88.90
Proposed approach
1D-LGP + ANN
It can be seen in Table 3 that the proposed model also achieves a high classification accuracy with 20 subjects.
4
Conclusion and Future Work
In this paper, a new framework with 1D-LGP and ANN has been proposed
for building an EEG based biometric identification system. The effectiveness
of the proposed framework has been tested with the publicly available EEG
dataset. The framework deals with multi-class classification problem of biometric
identification. The ANN classifier was used to carry out the classification process.
The experimental results are compared with some of the existing techniques in
the literature. From these experimental results, it could be interpreted that the
proposed model is effective for developing an EEG based biometric identification
system. In future, the applicability of the proposed model can also be tested in
different fields.
Acknowledgments. The authors would like to thank Henri Begleiter, Neurodynamics
Laboratory, State University of New York Health Center, Brooklyn, for the dataset used
in this research.
References
1. Del Pozo-Banos, M., Alonso, J.B., Ticay-Rivas, J.R., Travieso, C.M.: Electroencephalogram subject identification: a review. Expert Syst. Appl. 41(15), 6537–6554
(2014)
2. Marcel, S., del R. Millán, J.: Person authentication using brainwaves (EEG) and
maximum a posteriori model adaptation. IEEE Trans. Pattern Anal. Mach. Intell.
29(4), 743–752 (2007)
3. Palaniappan, R., Mandic, D.P.: Biometrics from brain electrical activity: a machine
learning approach. IEEE Trans. Pattern Anal. Mach. Intell. 29(4), 738–742 (2007)
Brain Signal Based Biometric Identification
535
4. Huang, X., Altahat, S., Tran, D., Shutao, L.: Human identification with electroencephalogram (EEG) for the future network security. In: Lopez, J., Huang, X.,
Sandhu, R. (eds.) NSS 2013. LNCS, vol. 7873, pp. 575–581. Springer, Heidelberg
(2013). doi:10.1007/978-3-642-38631-2 42
5. Yeom, S.-K., Suk, H.-I., Lee, S.-W.: Person authentication from neural activity of
face-specific visual self-representation. Pattern Recogn. 46(4), 1159–1169 (2013)
6. Yang, S., Deravi, F.: Novel HHT-based features for biometric identification using
EEG signals. In: ICPR, pp. 1922–1927 (2014)
7. Del Pozo-Banos, M., Travieso, C.M., Weidemann, C.T., Alonso, J.B.: EEG biometric identification: a thorough exploration of the time-frequency domain. J. Neural
Eng. 12(5), 056019 (2015)
8. Fraschini, M., Hillebrand, A., Demuru, M., Didaci, L., Marcialis, G.L.: An EEGbased biometric system using eigenvector centrality in resting state brain networks.
IEEE Sig. Process. Lett. 22(6), 666–670 (2015)
9. Maiorana, E., La Rocca, D., Campisi, P.: Eigenbrains and eigentensorbrains: parsimonious bases for EEG biometrics. Neurocomputing 171, 638–648 (2016)
10. Rodrigues, D., Silva, G.F.A., Papa, J.P., Marana, A.N., Yang, X.-S.: EEG-based
person identification through binary flower pollination algorithm. Expert Syst.
Appl. 62, 81–90 (2016)
11. Sharma, P.K., Vaish, A.: Individual identification based on neuro-signal using
motor movement and imaginary cognitive process. Optik Int. J. Light Electron
Optics 127(4), 2143–2148 (2016)
12. Ahonen, T., Hadid, A., Pietikainen, M.: Face description with local binary patterns:
application to face recognition. IEEE Trans. Pattern Anal. Mach. Intel. 28(12),
2037–2041 (2006)
13. Chatlani, N., Soraghan, J.J.: Local binary patterns for 1-D signal processing. In:
18th European Signal Processing Conference, EUSIPCO-2010, pp. 95–99 (2010)
14. Kaya, Y., Uyar, M., Tekin, R., Yıldırım, S.: 1D-local binary pattern based feature
extraction for classification of epileptic EEG signals. Appl. Math. Comput. 243,
209–219 (2014)
15. Sunil Kumar, T., Kanhangad, V., Pachori, R.B.: Classification of seizure and
seizure-free EEG signals using local binary patterns. Biomed. Sig. Process. Control
15, 33–40 (2015)
16. Jaiswal, A.K., Banka, H.: Local pattern transformation based feature extraction
techniques for classification of epileptic EEG signals. Biomed. Sig. Process. Control
34, 81–92 (2017)
17. Jun, B., Kim, D.: Robust face detection using local gradient patterns and evidence
accumulation. Pattern Recogn. 45(9), 3304–3316 (2012)
18. Jun, B., Choi, I., Kim, D.: Local transform features and hybridization for accurate
face and human detection. IEEE Trans. Pattern Anal. Mach. Intel. 35(6), 1423–
1436 (2013)
19. Huang, S.-C., Huang, Y.-F.: Bounds on the number of hidden neurons in multilayer
perceptrons. IEEE Trans. Neural Netw. 2(1), 47–55 (1991)
20. Abdullah, M.K., Subari, K.S., Loong, J.L.C., Ahmad, N.N.: Analysis of the EEG
signal for a practical biometric system. World Acad. Sci. Eng. Technol. 68, 1123–
1127 (2010)
21. Zhao, Q., Peng, H., Hu, B., Liu, Q., Liu, L., Qi, Y.B., Li, L.: Improving individual
identification in security check with an EEG based biometric solution. In: Yao, Y.,
Sun, R., Poggio, T., Liu, J., Zhong, N., Huang, J. (eds.) BI 2010. LNCS, vol. 6334,
pp. 145–155. Springer, Heidelberg (2010). doi:10.1007/978-3-642-15314-3 14
536
A.K. Jaiswal and H. Banka
22. Ma, L., Minett, J.W., Blu, T., Wang, W.S.Y.: Resting state EEG-based biometrics
for individual identification using convolutional neural networks. In: 2015 37th
Annual International Conference of the IEEE Engineering in Medicine and Biology
Society (EMBC), pp. 2848–2851. IEEE (2015)
23. Bai, Y., Zhang, Z., Ming, D.: Feature selection and channel optimization for biometric identification based on visual evoked potentials. In: 2014 19th International
Conference on Digital Signal Processing, pp. 772–776. IEEE (2014)
24. Maiorana, E., Rocca, D.L., Campisi, P.: Eigenbrains and eigentensorbrains: parsimonious bases for EEG biometrics. Neurocomputing 171, 638–648 (2016)
25. Kumari Sharma, P., Vaish, A.: Individual identification based on neuro-signal using
motor movement and imaginary cognitive process. Optik Int. J. Light Electron
Optics 127(4), 2143–2148 (2016)
26. Kumari, P., Vaish, A.: Brainwave based user identification system: a pilot study in
robotics environment. Robot. Autonomous Syst. 65, 15–23 (2015)
Документ
Категория
Без категории
Просмотров
5
Размер файла
733 Кб
Теги
978, 6427, 981
1/--страниц
Пожаловаться на содержимое документа