Вы находитесь на странице: 1из 6

(IJCNS) International Journal of Computer and Network Security, 15

Vol. 2, No. 1, January 2010

Improving Computing Efficiency of a Wavelet


Method Using ECG as a Biometric Modality
Jianchu Yao1 and Yongbo Wan2
1
Department of Engineering, East Carolina University
East 5th Street, Greenville, NC 27858
yaoj@ecu.edu
2
Department of Engineering, East Carolina University
East 5th Street, Greenville, NC 27858
wany@ ecu.edu

Abstract: Used extensively in cardiac diagnosis for decades, caused by ionic electrical signals. It has been demonstrated
electrocardiogram (ECG) has recently been investigated as a initially that ECG signals present a unique pattern of each
new biometric modality. Existing research reveals that ECG person’s heart properties and may be utilized for individual
identification using wavelet coefficients as the “statistic identification [3, 5-8].
attributes” outperforms other classification algorithms such as
correlation analysis. When wavelet coefficients are used, the R
classification process requires significant computation power,
which may pose an issue in circumstances where the device’s
computation capacity is limited. The paper examines the effect
T
on identification accuracy when only a subset of the wavelet
coefficients is used to calculate the signal difference/similarity P
measures. A Birge-Massart strategy is employed to select the
Q
coefficient subset. Results demonstrated that classification with S
the wavelet coefficient subset can achieve comparable
performance as when the entire coefficient structure is utilized.
The substantial reduction of the number of involved wavelet Figure 1. An ECG trace cycle.
coefficients improves real-time performance and eases
ECG as a new biometric trait will be useful for those
computation requirements so that this new identification
modality can potentially be implemented on portable devices. applications where ECG data are already collected and/or
stored in the system because, in these applications,
Keywords: Accuracy rate, computing efficiency, identification does not require extra equipment or data
electrocardiogram (ECG), human identification, wavelet collection, making prompt identification possible without
coefficients. increasing cost. One particular area that can benefit from
ECG identification is body sensor networks [7, 10], which
will be widely employed in future telemedicine and
1. Introduction and Background
homecare systems. These sensor networks monitor the
Since the “9/11” terrorism attack on the United States, patient’s status in an autonomous manner and require the
interest in biometric technologies has surged due to growing networks to be self-descriptive. In other words, the system
security concerns from the public and the government. should be able to register the owner of the signal
Biometric technologies can be generally classified into two automatically and store the collected data to correct
categories: physiological characteristics (e.g., fingerprint, electronic medical records. Identifying users with ECG
facial pattern, hand vein structure) and behavioral signals (instead of conventional biometrics that require
characteristics (e.g., gaits, keystrokes) [1]. As the additional data collection) simplifies the configuration and
applications of existing bioidentification technology are lowers the cost of such systems. ECG can also serve as a
wide spreading, new biometric modalities are being convenient identification tool in the battle field: when
explored to address these issues [2-4]. These new traits information about injured soldiers is sent back to hospital
include electrocardiogram (ECG) [3], facilities, their ECG signals can be conveniently used as
photoplethysmography (PPG) [2], and identifiers. Other applications may include identifying
electroencephalograph (EEG) [4]. Among these, ECG has patients from natural disasters, fire sites, terrorism attacks,
gained substantial attention during the last decade [3, 5-8]. etc. It should be noted that recent advancements of
A typical ECG wave cycle, as shown in Figure 1, consists of noncontact, miniature, and low-cost ECG devices should
a P wave, a QRS complex, and a T wave. Largely, the P further make collection of ECG signals convenient and
wave represents the depolarization of the myocytes affordable for non-hospital settings [11, 12].
spreading from the sinoartrial (SA) node to the Identification with ECG signals has been studied in both
atrioventricular (AV) node; the QRS complex represents the time and frequency domains [3, 5, 6, 13, 14]. The time
ventricular depolarization, and the T wave reflects the domain method [3] extracts time intervals and magnitudes
ventricular repolarization [9]. ECG waveforms, governed between ECG peaks and troughs as statistical classification
by the balance of sympathetic and parasympathetic heart attributes. While using temporal distances and magnitudes
system signals, represent the myocytes’ conductive activities for identification appears intuitive, the acquisition of these
16 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 1, January 2010
attributes might not always be easy for ECG waveforms identification with MATLAB scripts using the wavelet
corrupted by noises. Furthermore, extraction of time approach. The contribution of wavelet coefficient subsets is
intervals as statistical attributes ignores the details of the analyzed based-on a signal difference/similarity measure
singular points and lacks a comprehensive exploitation of defined similar to that in [5].
information contained in the ECG waveform. For example,
2.1 Data Acquisition and Signal Preparation
the QRS complex shape of different subjects may vary
substantially, yet the Q-R and R-S intervals remain the same In the experiment, a three-lead ECG sensor employing a low
(refer to Figure 1). The time interval method does not power precise instrumentation amplifier [16] was built to
account for these informative variations. collect signals. The three Ag/AgCl electrodes were attached
Alternatively, wavelet transform decomposes time series to both wrists and the left ankle of a subject to collect the
signals into spectral and temporal dimensions, offering sensed electrical cardiac signals to a conditioning circuit
inclusive insights about energy distribution of the signal through nickel plated brass snap type connectors. These
along the two axes. Wavelet decomposition has been analog ECG signals, after being amplified and initially
employed recently to develop classification algorithms [5, conditioned by the circuits, were acquired by a LabVIEW
7]. In a wavelet transform, a signal is decomposed into program running on a personal computer with a National
Instrument multifunction input/output card (DAQCardTM-
a combination of a set of basis wavelet functions , 6062E). The DAQCardTM-6062E data acquisition card is
where is the scale factor and is the translation equipped with sixteen 12-bit analog-digital conversion
parameter. The result of the wavelet transform is a structure channels which can sample at a rate up to 500 KS/s. Since
of wavelet coefficient . That is, no filtering function is provided by the conditioning circuit,
ECG data were first acquired at a rate of 10 KS/s and high
(1)
frequency noises were then filtered out by taking average of
Signal analysis with wavelet decomposition eliminates the every 50 samples, resulting in a sampling rate of 200 S/s. A
need for time interval extraction, as recognized by Chan, et LabVIEW interface displayed ECG signals real-time so that
al. and Yao, et al. [5, 7]. The former proposed a wavelet only data when the subjects were in stable condition would
distance approach for ECG identification, where the sum of be collected. Experimental parameters, such as sampling
normalized differences between wavelet coefficients was rate, record duration, and storage paths could be adjusted
used as the classification score to evaluate the similarity through the interface.
between two ECG signals. The research showed that the
wavelet approach outperforms the time domain approach
that uses correlation coefficient to measure difference
between signals.
Interested in innovative use of vital signs as biometrics [2,
7, 15], the authors evaluated the effectiveness of the wavelet
classification approach [7] and soon realized a computation
issue involved in this approach: the existing wavelet
distance algorithm did not distinguish the coefficients
containing salient information from those less informative
or even noise-corrupted ones. That is, it treated the
coefficients at all scales equally when calculating the
wavelet distances. Given that hundreds of coefficients are
obtained from an ECG wavelet transformation, the Figure 2. An "in-house" ECG sensor for data collection.
computation involved may exceed the capability of the
wearable/portable platforms for those potential user After appropriate internal research board approval,
scenarios mentioned earlier (i.e., body sensor network, etc). volunteers were recruited for this study. The subjects were
This paper hence investigates two issues: (1) whether a comprised of 26 males and 4 females, with ages ranging
subset of wavelet coefficients works equivalently as the from 18 to 51 years. A total of 121 datasets from 30
entire structure of coefficients in determining the distance subjects were collected. Each subject participated in
measure; and (2) if a coefficient subset can obtain multiple data collection sessions. Session intervals varied
comparable classification results as the entire structure, from several hours (morning and afternoon sessions) to a
determine the saved computation consumption. few weeks. Each session lasted around five minutes: after
the subject rested in a chair for two minutes, the ECG leads
2. Methods were placed; two minutes of consistent data were then
This section describes the steps to investigate the two acquired and stored. Each subject participated in at least
critical issues for computation efficiency improvement. two sessions. The maximum of sessions participated in
These steps include data collection, signal pre-processing, were by two subjections who participated in five sessions.
subject classification, coefficient selection, and computation In other words, for the ith subject, the number of collected
experiments. This study employed an “in-house” device to datasets : and where i is the subject number.
collect ECG signals as show in Figure 2. Data acquired The stored raw signals were pre-processed with a
through a LabVIEW program were processed for MATLAB program. Two major noise sources—signal drifts
(IJCNS) International Journal of Computer and Network Security, 17
Vol. 2, No. 1, January 2010
at around 0.06 Hz and spikes at 60 Hz—were first filtered select constant for normalization purpose. To obtain the
with hard thresholding after a scale 12 Daubechies’s db6 DDWC measure, a scale 6, bior1.1 wavelet transform [18]
[17] wavelet transform to all the heart beat cycles. Detail was applied to the pre-processed ECG waveforms, resulting
wavelet parts at scales 2, 3, and 4 were reconstructed so that in coefficient structures of 256 elements. The ‘Bior1.1’
the R peaks (refer to Figure 1) could be located as the basis function belongs to the Biorthogonal Wavelet Pairs
fiducial point to identify ECG cycles. Identified ECG cycles wavelet family; the orthogonal discrete wavelet transform
were interpolated to a pre-defined length for the functions have excellent localization properties both in time
convenience of future steps. Sixty consistent heartbeat and frequency domains [19] and coefficients obtained
cycles from each dataset were selected and normalized to the contain distinctive information. Note that the basis function
range of [-1, 1]. In this step, the consistency of data was chosen here is different from that in [5], where a db3
examined with Euclidean distances of each cycle to the function was used.
mean of all cycles.
2.3 Wavelet Coefficients with Salient
A wavelet transform was applied to these time series
Information
signals and a set of coefficients were calculated for each of
the sixty heart beats. Depending on the wavelet transform The magnitudes of the wavelet coefficients decomposed
(continuous or discrete) applied, the number of coefficients from an ECG signal differ substantially from scale to scale.
vary (more information is provided when the specific Table I illustrates the ranges of coefficients from a scale 6
measures are introduced). Nevertheless, for each of the 121 discrete wavelet transform. The magnitudes of the detailed
ECG datasets acquired earlier, wavelet coefficients for their coefficients at scale 3 (from -6.89 to 0.52) were significantly
sixty heart beats were saved. From this point on, all the greater than those at scale 1 (in range of -0.27 to -0.297).
operations were performed on these wavelet coefficients. Considering all the coefficients in such a broad range, one
They served as the “statistical features” from which would conclude that the major components with larger
classification scores were calculated and subsequent magnitudes contribute more in discriminating one subject
classification decisions were made. from another. On the other hand, since ECG waveforms are
Out of the Ni coefficient sets (each corresponds to sixty similar, their major components must be largely alike.
heartbeats), one coefficient set is enrolled in the database, Therefore, one might argue that it should be those minor (or
creating a database of 30 coefficient sets. The other Ni-1 subtle) components that really make the difference in
coefficient sets of the ith subject were used to test the distinguishing subjects. A more general hypothesis
classification accuracy; altogether there are 91 (that is, 121- warrants investigation: a subset of the coefficients from the
30=91) coefficient sets available for classification test. wavelet coefficient structure can obtain equivalent
identification results as when the entire structure is used.
2.2 Distance of Discrete Wavelet Coefficients In this study, subsets of coefficients from the
(DDWC) decomposition structure were selected using a MATLAB
The classification method in the experiment finds the function ‘wdcbm(C, L, ALPHA, M) [20]. This function
distances (D) from the to-be-tested coefficient set employs an adaptive coefficient selection rule based on the
to those enrolled coefficient Birge-Massart (BM) strategy [21]. The input arguments of
sets in the database and uses these the function C and L describe the wavelet coefficient
structure; and ALPHA and M are constants and control the
distances as the quantitative measure of signal number of selected coefficients. The function returns the
difference/similarity. After all the distances were compared, number of selected coefficients and the scale-dependent
is classified to the closest enrolled subject . i.e., thresholds. Once the coefficient subset was selected, the
unknown coefficient set distance measure ( , distance of main discrete
wavelet coefficients) was calculated using the equation
, where (2) below:
To classify a subject, a “distance” measure needs to be
defined to represent the level of similarity/difference
between unknown wavelet coefficient set and the enrolled
coefficient sets. Here a distance of discrete wavelet (4)
coefficients (DDWC) is defined similar to [5], where it was
referred as WDIST in [5]. For an unknown subject, its Where is the coefficient subset selected with BM
DDWC to the n-th enrolled subject is given by: strategy; all other notations follow those in (3), except that
the denominator takes a slightly different form.
(3) Identification performance of different coefficient subset
sizes, ranging from 10 to 100 with an increment of 10, was
where is the th wavelet coefficient at the th scale of examined with the measure. For each specific
the unknown coefficient set; is the th wavelet number of selected coefficients, a cross-validation of ten
trials was utilized to estimate the relationship of accuracy
coefficient at the th scale of the enrolled coefficient set; is
rate ( ) versus the coefficient numbers.
the number scales of the wavelet transform; is the
number of coefficients at a specific scale. T.H. is a pre-
18 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 1, January 2010

Table 1: Magnitudes of Wavelet Coefficients Vary from Scale to Scale


Approximation Detailed Detailed Detailed Detailed Detailed Detailed
Components
at scale 6 at scale 6 at scale 5 at scale 4 at scale 3 at scale 2 at scale 1
Number of
4 4 8 16 32 64 128
Coefficients
Range of -7.17 -0.05 -2.09 -0.40 -6.89 -0.79 -0.27
Magnitudes ~-5.78 ~1.68 ~0.98 ~2.28 ~0.52 ~0.59 ~0.297

And the standard deviations from these trials were larger number of subjects, more detailed, high-frequency
compared to find the performance trend as the coefficient information is required.
number increases. Additionally, possible variations of this
Figure 3 also shows that, when the number of selected
relationship with the number of subjects changes was also
coefficients is 40, there appears to be an unusual drop in
investigated by analyzing results from five cases, each of
performance (i.e., lower performance than when the number
which include data from 5, 10, 15, 20, or 25 subjects.
of selected coefficients is 30 or 50). A closer examination
2.4 Computing Time versus Number of on coefficients at different wavelet scales was taken to find
Coefficients Used for Classification out the causes. The coefficients after the first thirty were
The significance of computation saving because of the selected primarily from scale 3 (frequency range of 12~25
reduction of the number of wavelet coefficients utilized for Hz). At first, only a small portion of coefficients (~5 out of
classification was studied by examining the execution time 32) at this frequency range were selected. Both magnitudes
of the classification algorithms. In the experiments, the and locations (along the time axis) of the selected
code was programmed in MATLAB R2007b and run on a coefficients change considerably and consequently introduce
personal computer with Microsoft XP Professional operating variations to the distance measure, causing performance
system. The computer CPU operates at a 3.4 GHz clock and deterioration. As the number of used coefficients increases
works with 1.0 GB physical memory and 2.0 GB virtual to 50, more scale 3 coefficients would be included, the
memory. Computing time for different tasks required to detailed components of the ECG signal vary less and
identify unknown datasets was recorded. These tasks classification performance resume.
include signal preprocessing (wavelet transform, filtering, 100

signal selection, and necessary BM coefficient selection) and


classification of the unknown dataset with reduced number 95

(out of 256) of coefficients. Note that the computing time to


90
obtain wavelet coefficients enrolled in the database is a one-
Accuracy Rate (%)

time cost and does not affect real-time identification.


85
Therefore, database enrollment preparation was not
included in the study.
80
5 Subjects
3. Results 75
10 Subjects
15 Subjects
20 Subjects
25 Subjects
Figure 3 depicts the identification performance when only a 70
10 20 30 40 50 60 70 80 90 100
subset of discrete wavelet coefficients selected with the Number of Coefficients in the Selected Subset
Birge-Massart strategy was employed for discrimination. It
can be seen from the figure that, for all the cases, the Figure. 3 The DDWC identification accuracy rate vs. the
identification rate starts to decrease noticeably if fewer than number of selected coefficients.
50 out of the 256 coefficients are used in the DDWC
calculation. It is also interpreted from this figure that, when
the number of subjects to be identified is small (5 or 10), the Figure 4 shows the normalized computing time saving as
performance starts to deteriorate noticeably with a small the number of wavelet coefficients used for classification is
subset (< 30) of the 256 coefficients. These coefficients reduced. As illustrated by the figure, computing time
selected with the Birge-Massart strategy are mostly the required for classification is proportional to the coefficient
wavelet approximation and low-frequency components (<12 number: when only the classification computation is
Hz) of the original signal. Other coefficients that provide concerned (“Classification Only”), using 50 coefficients
more detailed information (frequency >12Hz) do not seem to takes about 45% of time as that required by using 256
improve the accuracy rate in these cases. coefficients, demonstrating a computation saving of 55%.
In large subject number (20 or 25) cases, however, a slight The preprocessing of the unknown dataset takes 0.4~0.5
performance improvement can be noticed when the number seconds. When this coefficient preparation was included,
of coefficients increased from 50 to 100. This small the overall computing time saves 22% when 50 out of 256
accuracy increase could be a result of a higher degree of coefficients are used as suggested by the “Prep Included”
freedom as a larger number of coefficients were involved in curve.
classification or it may indicate that, in order to identify a
(IJCNS) International Journal of Computer and Network Security, 19
Vol. 2, No. 1, January 2010

The reduction in the number of involved coefficients can


100
bring considerable computation saving (>20%) to the
90 distance classification method used in this paper. This
Normalized Required Time (%)
reduction in the number of used coefficients is expected to
80 bring more computation benefits for biometric systems that
use intelligent classification/discrimination methods (e.g.,
70 artificial neural network), whose computation efficiency and
performance are sensitive to the size of input vectors. The
60
computation improvement will make ECG better suited to
50
those applications where computation power is limited, such
as those body area sensor networks that are supposed to be
40 Classfication Only ubiquitous in the healthcare sector in the near future.
Prep Included

30
0 50 100 150 200 250 300 5. Conclusions
Number of Coefficients Used for Classification
This paper investigated a possible ways of improving
Figure 4. Computing time saving with the reduced number computing efficiency of a wavelet approach for human
of coefficients used for classification. identification with ECG signals. The project collected data
with an “in-house” ECG sensor. Wavelet coefficient sets
4. Discussion after pre-process served as the statistical features for
classification; the linear classification method used a
Our experiment results with the wavelet approach obtained discrete wavelet distance measure to describe the similarity
an identification accuracy rate of higher than 80%. These between these statistical features. The paper examined the
results are generally in agreement with those from identification performance as a function of the number of
previously published work and demonstrate the promise of wavelet coefficients included for calculating the distance
ECG waveforms as a biometric modality, confirming the measure. Results from the research shows that it is possible
potential of using ECG signal as a biometric modality for to use only a portion of the wavelet coefficients in the
security purposes. Compared to many existing biometric classification process to gain the best identification
traits, ECG is promising to address several concerns: (1) the performance. The reduction of the number of coefficients
collection of ECG signals is not expected to change with will improve identification efficiency and performance,
environmental conditions as those conventional traits particularly in those methods whose computation
collected from body surfaces. In other words, the operation requirements are sensitive to the number of elements
of systems with ECG identification may be less susceptive to contained in the input vectors.
external noises and demonstrate more stable performance;
(2) ECG signals, originating from the deep interior of the
References
human body, reflect physiological activities of the heart and
other parts of the cardiac system. It is thus extremely hard [1] A. K. Jain, A. Ross, and S. Prabhakar, "An
for one to spoof any legitimately enrolled measurements introduction to biometric recognition," IEEE
with this trait; and (3) the presence of ‘live’ subjects Transactions on Circuits and Systems for Video
guarantees successful collection of the biometric Technology, vol. 14, pp. 4- 20, 2004.
characteristics and, therefore, eliminates possible failure to [2] J. Yao, X. Sun, and Y. Wan, "A pilot study on using
enroll. derivatives of photoplethysmographic signals as a
Figure 3 illustrates that a small number of coefficients biometric identifier," in the 29th International
(20~30) selected by the Birge-Massart strategy offers very Conference of the IEEE Engineering in Medicine and
good identification results when the number of to-be-tested SFGBM, Lyon, France, 2007.
subjects is small. These coefficients depict the [3] S. Israel, J. M. Irvine, A. Cheng, M. D.Wiederhold,
approximation and low-frequency components of the ECG and B. K.Wiederhold, "ECG to identify individuals,"
waveform. This approximation, without much detailed Pattern Recognition Society, vol. 38, pp. 133-142,
information, suffices when the subject size is not large. 2005.
When the subject population is larger, on the other hand, [4] S. Marcel and J. D. R. Millan, "Person authentication
more wavelet coefficients must be included to obtain using brainwaves (EEG) and maximum a posteriori
reasonable results because additional information provided model adaptation," IEEE Transactions on Information
by these coefficients helps to distinguish subjects compared. Technology in Biomedicine Pattern Analysis and
The performance improves initially as the number of Machine Intelligence, vol. 29, pp. 743 - 752, 2007.
selected coefficients increases; however, this improvement [5] A. D. C. Chan, M. M. Hamdy, A. Badre, and V.
does not extend to the entire coefficient structure. The Badee, "Wavelet distance measure for person
performance achieves its maximum when the coefficient identification using electrocardiograms," IEEE
number reaches approximately 80. Regardless of smaller or Transactions on Instrumentation and Measurement,
larger subject sizes, it is shown that identification with a vol. 57, FEBRUARY 2008.
subset of wavelet coefficients can obtain equivalent [6] L. Biel, O. Pettersson, L. Philipson, and P. Wide,
performance as that when all the coefficients are considered. "ECG analysis: A new approach in human
20 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 1, January 2010

identification," IEEE Transactions on Instrumentation Authors Profile


and Measurement, vol. 50, pp. 808–812, 2001.
[7] J. Yao and Y. Wan, "A wavelet method for biometric
identification using wearable ECG sensors," in the 5th Jianchu Yao received the Ph.D.
International Workshop on Wearable and Implantable degree in electrical engineering from
Body Sensor Networks (BSN 2008), Hong Kong, 2008. Kansas State University (KSU),
[8] Z. Zhang and D. Wei, "A new ECG identification Manhattan, KS, in 2005. He is
method using Bayes' Theorem," in TENCON IEEE currently an Assistant Professor in the
Region 10 Conference, Hong Kong, 2006, pp. 1 - 4. Department of Engineering, East
Carolina University (ECU),
[9] J. G. Webster, Medical instrumentation: Application
Greenville, NC, where he develops
and design: Wiley, 1998. and directs the Bio-sensors Laboratory
[10] C. C. Y. Poon, Y.-T. Zhang, and S.-D. Bao, "A novel with research areas focused on new medical sensors and innovative
biometrics method to secure wireless body area sensor applications of vital signs. From 1995 to 2001, he served as a
networks for telemedicine and m-health," IEEE Controls Engineer at two research institutions in China. Dr. Yao
Communications Magazine, vol. 44, pp. 73 - 81, 2006. is a senior member of IEEE and a member of the American Society
[11] M. Steffen, A. Aleksandrowicz, and S. Leonhardt, of Engineering Education.
"Mobile noncontact monitoring of heart and lung
activity," IEEE Transactions on Biomedical Circuits Yongbo Wan received the B.S.E.E. and
and Systems, vol. 1, pp. 250 - 257, 2007. M.S.E.E. degrees in electrical engineering
from Shaanxi University of Science &
[12] T. J. Sullivan, S. R. Deiss, and G. Cauwenberghs, "A Technology (SUST), Xi’an, Shaanxi,
low-noise, non-contact EEG/ECG sensor," in IEEE China, in 2001 and 2006, respectively.
Biomedical Circuits and Systems Conference, 2007, From 2007 to 2008, he worked as an
Montreal, Canada, 2007, pp. 154 - 157. exchange scholar in the Department of
[13] Z. Zhang and D. Wei, "A New ECG Identification Engineering at East Carolina University,
Method Using Bayes' Teorem," in TENCON IEEE NC, USA. He is currently a doctoral student
Region 10 Conference, 14-17 Nov. 2006, pp. 1 - 4. at Oklahoma State University. His research
[14] M. Kyoso and A. Uchiyama, "Development of an ECG interests include applications of embedded systems, wireless data
identification system," Proc. 23rd IEEE Eng. Med. transmission, real-time computing, medical electronics, and digital
signal processing.
Biol. Conf., vol. 4, pp. 3721-3723, 2001.
[15] Y. Wan and J. Yao, "A neural network to identify
human subjects with electrocardiogram signals," in
The World Congress on Engineering and Computer
Science (WCECS) 2008 of International Association of
Engineers (IAENG), San Francisco, CA, 2008.
[16] Linear_Technology, "LT1167-Single Resistor Gain
Programmable, Precision Instrumentation Amplifier,"
Linear Technology Corporation Datasheet, 1998.
[17] I. Daubechies, Ten lectures on wavelets: Society for
Industrial and Applied Mathematics, Philadelphia,
PA, 1992.
[18] S. Mallat, A wavelet tour of signal processing, 2 ed.:
Academic Press, 1999.
[19] G. K. Kharate, A. A. Ghatol, and P. P. Rege,
"Selection of mother wavelet for image compression
on basis of image," in IEEE - ICSCN 2007 Signal
Processing, Communications and Networking, Anna
University, Chennai, India, 2007, pp. 281-285.
[20] "MATLAB Documentation: Wavelet tool box,"
http://www.mathworks.com/access/helpdesk/help/toolb
ox/wavelet/index.html?/access/helpdesk/help/toolbox/
wavelet/ref_open.html&http://www.mathworks.com/pr
oducts/wavelet/.
[21] L. Birgé and P. Massart, "From model selection to
adaptive estimation," D. Pollard (ed), Festchrift for L.
Le Cam, Springer., pp.55-88, 1997.

Вам также может понравиться