Content uploaded by Abdulbasit Alazzawi
Author content
All content in this area was uploaded by Abdulbasit Alazzawi on May 06, 2018
Content may be subject to copyright.
IJCSNS International Journal of Computer Science and Network Security, VOL.18 No.3, March 2018
169
Manuscript received March 5, 2018
Manuscript revised March 20, 2018
Summary
Face recognition system has been widely utilized for various
sensitive applications such as Airport gates, special monitoring,
and tracking system. The performance of most face recognition
systems would significantly decrease if there were several
variations in the illumination of dataset images. In this paper the
proposed a new algorithm based on a combination of edge
detection operators, features extractors and artificial neural
network ANN as a classifier. The Second based on Laplacian
comprise Zero cross, Laplacian of gaussian LOG, and Canny
edge detection filters. A segmentation process is used to segment
each image to equaled size blocks treats face edge pixels
precisely. A new features extractor technique based on Linear
Regression Slope SLP with discrete wavelet transformation
(DWT) and principle components analysis PCA used for
features extraction. ANN used for the data set classification and
all results obtained evaluated. We tried a combination of various
techniques like (Zero cross, DWT, SLP-PCA, ANN),(LOG,
DWT, SLP-PCA, ANN),(Canny, DWT, SLP-PCA, ANN ). The
proposed method is examined and evaluated with different face
datasets using ANN classifier. The experimental results were
displaying the superiority of the proposed algorithm over the
algorithms that used the state-of-art techniques where the
combinations (Zero cross, SLP, ANN) gained the best results and
could outperform all the other algorithms.
Key words:
Face Recognition, SLP, PCA, Neural Network, ANN
1. Introduction
Since from last decade, face recognition system has had
vital effects in daily life, especially for security objectives.
The system is the strong domain for human being
authentication process in various authentication
applications. The systems have been widely utilized for
various sensitive applications such as Airport gates,
special monitoring, and tracking system, criminal
distinguishing, broad checkpoints and many other
applications. Face recognition used for identity
identification or verification process. Face identification
process is nothing except one to one matching problem.
Meanwhile, face verification is more complicated because
it is one to many matching problems. It is utilized to
identify the tested face against huge face images that saved
in the database. Many approaches, and techniques used in
face recognition process. Generally, we can divide it into
two categories. First, one is to data reduction and features
extraction such as holistic methods that include,
eigenvectors based on PC [1], independent component
analysis technique (ICA) [2], linear discriminant analysis
(LDA) technique [3], and kernel LDA [4].The second one
is utilized as a classifier to find the faces features which
are most likely to be looking for such as Neural Network-
based approaches [5] Support vector machine, and nearest
distance [6]. Most of these approaches and techniques
have trade-offs such as time for features extractions, the
response time for training dataset, lack to update training
dataset and hardware requirements.
In this study, we chose a gradient, Laplacian-based
operators for edges detection, principal components
analysis PCA, slope based method SLP for feature
extraction, and ANN for a classification.
The edges are the crucial property that can be used for data
reduction and feature detection. Data reduction comes
from excluding all image data except edge. These data can
be used to find face objects features. Many operators used
for edge detection. Generally, we used second-order
derivative operators that includes Zero cross, LOG and
Canny filters based on Laplacian operator. All these filters
have weaknesses during edge detection process such as
weakness in noise depressing, find out some type of edge
but not all and lack edge shape. Therefore, we connect
those edge detection filters with discrete wavelet
transformation DWT for edge getting optimization.
DWT is very well known in image decomposition by
separation original data set image to approximation image
and three detail images (horizontal, diagonal, and vertical
image. Wavelet transformation used for dimensionality
reduction. Also, it is used for time-space frequency
analysis. Wavelet transformation provides time-frequency
analysis for one or two-dimensional signal which is
practically powerful in image analysis, computer vision,
Evaluation of Face Recognition Techniques Using 2nd Order
Derivative and New Feature Extraction Method Based on Linear
Regression Slope
Abdulbasit Alazzawi†,††, Osman N. Ucan†, Oguz Bayat†
† Altinbas University, Department of Electrical and Electronics Engineering, Istanbul, Turkey
†† Diyala University, College of Science, Diyala, Iraq
IJCSNS International Journal of Computer Science and Network Security, VOL.18 No.3, March 2018
170
pattern recognition. Unlike Fourier transformation that
used only for time analysis signals [7].
PCA is powerful and long-term studies technique applied
for feature extraction from face images by creating a face
feature space. PCA has another advantage that is low
computation time because it depends on modified
covariance matrix to find Eigenfaces. Whoever, because of
being linear features extraction, PCA is less effective
especially when nonlinearity issue take place in the
underlying relationships [8].
Slope base method SLP is a new method used for features
extraction and became dependable in IEEE based on [9].
The algorithm of SLP depends on linear regression slope.
The neural network is well known parallel processing units
(neurons) container. The main challenges of the neural
network are training dataset such that output same efficient
rate in the testing phase (generalization).
The experimental analysis of proposed framework is done
using MATLAB image processing toolbox with three
different face datasets such as [10-13]. This dataset
prepared under the different conditions and having various
image qualities, illumination conditions, lightning effects,
occlusions etc. Also, this study, we used 23 subjects
classes in the face BIO ID database [14].
The sections of the paper are organized as follows: in
sections 2, we present edge detection. Section 3 presents
features extraction methods (DWT, SLP, PCA). These
methods are discrete wavelet transformation DWT,
eigenfaces based on principal components analysis PCA
and slope based method SLP. Section3 introduces
segmentation process based on divide and conquer
principle. In section4, we introduce neural network as a
classification method. Section 5 shows the experimental
results and section 6 produces conclusions.
2. Methodology
The following diagram explains the steps of the proposed
system .as we can see, the procedure of process starts with
the select dataset for training ANN. Whoever, before reach
to last step (training ANN) we performed some steps,
these steps include applying DWT on the training data set
images and select a LOW LOW image. performing edge
detection operators on the approximating image is the
second step. Many edge detection operators were used
since each of which covers some types of edges, therefore
we used multi filters to cover all the types of edges.
segmentation process is used to segment the image that
came from edge detection process. Each image is
converted to many eaqeulad size blocks. Then, form these
blocks, features are extracted using new suggested method
with SLP name that depends on linear regression slope .to
measure the efficiency of the SLP state of art are used,
eigenface based on PCA. As a classifier, efficient Artifice
neural network is used.
3. Edge Detection Methods
3.1 Zeroross Operator
The zero-crossing edge detector figures out zero-crossing
regions in images where the values of the Laplacian cross
through zero regions. According to this operator, the edge
point pixels are mostly represented by points that have a
value of the Laplacian passes through zero which means
that in these points the Laplacian is changing its sign. For
example, in face recognition algorithms, the points where
the intensity of the face image edges changes suddenly is
considered as edge points candidates. The scientists
recommend dealing with this filter as feature extractor
operator since the output of the zero-crossing algorithm is
usually a binary image with one-pixel thickness [15-18].
Fig.1 Main Diagram of the Proposed System
3.2 LOG Operator
The Laplacian of Gaussian (LOG) algorithm is proposed
to solve the noise sensitivity of the regular Laplacian filter.
The steps of this algorithm start with removing noise by
applying image smoothing using a Gaussian blurring
IJCSNS International Journal of Computer Science and Network Security, VOL.18 No.3, March 2018
171
technique to obtain the best performance. After applying
the Laplacian method, change in sign from positive to
negative or vice-versa for any point (changes in sign)
represents an edge. This algorithm supposes that important
cross points almost be at edges in the images - i.e. where
the intensity values of the images regions changes sharply.
However, they also could occur at places that are not as
easy to allocate with image edges because some of these
regions are very like each other, especially in grayscale
images. The lines that are detected by the LOG procedure
have one-pixel thickness. The Laplacian L (x, y) of an
image with pixel intensity values I (x, y) is given by:
(,)=
+
(1)
This an important pre-processing operation decreases the
unwanted frequency noise prior to the differentiation
operation. The2-D$ LOG function centered on zero and
with Gaussian standard deviation has the form:
LOG (x, y) =(
*((1-
)* (e
)) ) (2)
the Gaussian smoothing filter and the Laplacian filter f and
then convolve this output result (hybrid filter) with the
digital image to reach the desired output [18].
3.3 Canny Operator
Canny edge detection operator uses five different and
sequential procedures to detect the edges; the first step is
smoothing the image to remove possible noise using a
Gaussian filter. The second step is computing gradient
values in vertical (Gx) and horizontal (Gy) directions
which is the same procedure that used in the Sobel
algorithm. Canny filter eliminates pixels that are not part
of image edges by using a non-maximum suppression
process. In the fourth step, threshold values are used to
determine the potential edges. In the last step of the Canny
algorithm, all edge pixels that have weak value or isolated
from other pixels are eliminated [19],[20]. For more
explanation, the mathematical model of the canny filter
includes the following equations.
G (m.n) = Gσ (m, n)
f (m, n) (3)
Where
=
e- (
) (4)
Compute gradient of g (m, n) using any of the gradient
operators to get the magnitude(M) and the direction () by
(,)=
(,)+
(,) (5)
and
(,)= (,)
(,) (6)
Where threshold is chosen so that all edge elements are
kept while most of the noise is suppressed.
4. Segmentation
One of the basic steps in the proposed system is the
segmentation of the image to equal size blocks. The idea
behind dividing segment image to sub-images is inspired
by the problem-solving technique. A popular technique
that used in advance pattern recognition algorithm and we
benefit from it is a divide and conquer technique [21] [22].
The dividing problem to many subproblems can help to
solve complex problems such as pattern recognition. In
this study, we divide the image into many sub-images for
two reasons. One is to find the edge values positions
precisely. Finding positions of the edge pixels is the basic
step for using SLP features extraction method. Second is
to separate noise pixels from edge pixels. Both of them we
tried are to simplify feature extraction process. The
number of segments parameter is a free number because
there is no standard value of the number of segments. I.e.,
we tried many scenarios (it is related to other parameters)
to select an optimal number of segments.
5. Features Extraction Methods
5.1 DWT Method
DWT is designed in a way such that we got good
frequency resolution for low-frequency components.2D
face signal translated into shifted and scaled levels
according to these two operations we obtained 4 sub
signals (sub-bands) viz; low -low LL, low high LH, high-
low HL, and high-high HH. In the proposal system, we
applied the multi-level discrete Wavelet transformation.
Low pass filter applied on row and columns of the image
to get LL subband. Low Low or approximation(A)
presents smaller scaled from the input image. Low pass
filter and high pass filter applied on the image to get the
three sub-bands, LH, HL, and HH. The most important
subband is low low(LL). We can construct all image from
this subband since this sub-signal shows a general trend of
pixel values. Also, if the details are small, we can ignore it
[23] [24] [25]. representation of this process on the image
as follows:
I (i, j) = (LL) + (LH + HL + HH) (7)
Or:
(,)=
+ (
+
+
) (8)
IJCSNS International Journal of Computer Science and Network Security, VOL.18 No.3, March 2018
172
this formal used for (n) times DWT decompositions. In our
study, we used:
n=21:23db (9)
-db is a Daubechies DWT.
we represented DWT decomposition levels that used in
this algorithm like:
(,)=
+
+
+
(10)
(,)=
+
+
+
(11)
(,)=
+
+
+
(12)
(,)=
+
+
+
(13)
5.2 Eigenface Method
An Eigenface-finding based on Principal Component
Analysis technique (PCA was also known as Karhunen
Love) is specialized for face image data. In principal
component analysis method, all dataset images are
recognized with feature vectors that are created from
projections of the training image to the basis in image
space. In general, PCA classifies dataset images according
to calculated distances among feature vectors. Typical
classifiers comprise nearest distance measure, Euclidean
distance, and nearest mean classification. Using PCA for
the Eigenfaces method, feature vectors identifying each
image is shown in figure (5), and these vectors can be
obtained as follows:
(1) Assume that we have (n) faces with (m) rows and
(m) column, as S = S1, S2,S...S T (14)
Represent the images in column vector space with
(m2 * 1) dimensions. According to the PCA
algorithm, we must calculate a mean face image (Q)
(common features for all dataset), from vectors by
the following Eq.
=
(15)
Where” Q” is the mean matrix (common properties
of faces).
(2) The second step is to subtract the main data matrix
from (S n), which can be represented as:
Ln = Sn − Q (16)
(3) Ln column vectors are gathered in the matrix:
R = [L1, L2, . . ., LN] with dimension (m2 * N) and
a covariance matrix C is formed as
=
= R.RT (17)
Because of high dimensionality and (considering
the linear algebra principle) the great computational
complexity of A AT multiplication, we use ATA
instead.
=
, C = RT.R (18)
(4) At this step, we calculate N eigenvalues (λp) and N
eigenvectors (VP) of C to form the eigenface
space. V = [v1, v2, . . ., vN] represents a matrix
including eigenvectors of (C) with a dimension of
(N * N). We can obtain eigenface space U=
[u1,u2...,uN]T by
U = V.RT (19)
All row vectors of (U) are eigenfaces of face
images in the training set. Face images with higher
eigenvalues have a greater contribution to the
eigenface space. For this reason, systems with low
computational capability sort eigenvectors of face
images according to their corresponding
eigenvalues in decreasing order and choose
eigenvectors to form a smaller eigenface space.
W = w1, w2, w3...,wN (20)
Matrix with dimension (N*N) includes (N) column
vectors corresponding to each face image in the
training set. These vectors are called feature vectors
and they represent each image specific
characteristics. (W) can be obtained
W = U. R (21)
After obtaining eigenface space and feature ectors,
we can compare a test image with the faces in the
training set by projecting a test image into faces
space as follows:
(a) ”ST” is a column vector which represents our test
image with (m2 1) dimension. At this stage, the
distance of test image from mean face image should
be calculated as LT column - vector,
LT = S T – Q (22)
(b) After calculating LT, we must project it on our
eigenface space in order to obtain its feature vectors
in the format of the column vector wT with
dimension (N *1),
wT = U.LT (23)
(c) find out which image in the training set resembles
our test image, we need to find the similarity of wT
to each (wi) in the matrix (W). Various classifiers
can be used at this stage. The technique we used is
the artificial neural network. [26-29].
5.3 Simple Linear Regression Slope-Based Method
To efficiently extract the best features from edge images, a
new feature extraction method that based on the slope of
the estimated curve of each small segment is used. The
proposed slope based feature extraction method (SLP)
produces vectors of features that can identify each face
successfully as shown in figure (6). The following steps
IJCSNS International Journal of Computer Science and Network Security, VOL.18 No.3, March 2018
173
explain in detail the process of extracting the features in
the training phase:
Fig. 2 SLP feature Extraction Method
Suppose that we have N face images (im1, im2,
im3...imN) with m rows and n columns, where N is the
number of images in the training set.
(1) Convert the face images (im1, im2, im3...imN) to
binary images using one of the six edge detection
filters that discussed in section 2. The binary
images denoted as (B1, B2, B3...BN).
(2) For each binary image Bi, divide it into (H L)
segments, where H is the number of rows and L is
the number of columns. The segments of image
face denoted by (Si1, Si2...Sik), where k is the
number of segments with (H L) dimension.
(3) For each segment Sij of each face image Bi, the
value of pixels with ones are considered as points
and a simple linear regression estimation method is
used to find the best line that fit these points.
Assume that the equation of the line is as
Y = m x + b (24)
Where b is the Y-intercept and m is the slope of the
estimated line.
(4) Compute the slope of the estimated line of each
segment Sij using the following equation which is
derived from Pearson’s correlation coefficient
formula and simple linear regression:
=( ()()())
(())() (25)
Where m is the slope of Sij segment, x is the x-coordinates,
y is the y-coordinates and n is the length of x, y vector in
the binary image. In our work, we select the linear
regression techniques because it is the most widely used
statistical technique to model a relationship between two
sets of variables and it is widely used to estimate the line
equation from a set of points. Slopes of segments are
collected in a matrix with dimension (H L)
corresponding to each face image in the training set to
create features vectors which will be the input to the ANN.
The usage of the simple linear regression techniques in our
steps makes the proposed algorithm works very well since
it generates very robust features that are not affecting by
outlier points or noise points that may exist in the binary
edge images.
6. Classification Process
The last step in the proposed system is training neural
network on the face images dataset. The input matrix of
the training dataset consists of the features vectors after we
applied all the preprocessing operations, dimensionality
reduction, and feature extraction. Steps of this algorithm
are explained as follows.
(1) First, prepare training data set.
(2) calculate the weighted sum of the input vectors (dot
product between the input vector and weights that
connect the input and the hidden layers) [30].
=
(26)
(3) The results that we obtain from the weighted sum
calculation is passed to the activation function to
normalize the output. In the proposed MLNN layers
we used the sigmoid function.
= (
()) (27)
(4) In the same way, we calculated the output of the
final layer by calculation of the weighted sum
between the hidden layer and an output layer.
=
(28)
= (
()) (29)
(5) After the output has been calculated, we calculate
the differences between actual output and the
desired output by using the minimum square error
between them.
(6) error=
(30)
∆Wli = αδ lHl (31)
δl = (Di − TrueNOuti) TrueNOuti (Hl −
TrueNOuti)
(32)
IJCSNS International Journal of Computer Science and Network Security, VOL.18 No.3, March 2018
174
7. Results and discussions
We performed face recognition based on two groups of
edge detection filters gradient and Laplacian-based
operators, each of which includes three filters. The second
order derivative operates includes LOG, Zero cross, and
Canny Filters. A common feature of the second group is,
all of them perform Laplacian process before they are
starting edge detection process. For the features extraction,
we used the eigenvectors-based method and suggested a
method, SLP. The ANN is the classifier that used for the
classification process. For DWT, PCA and SLP feature
extraction methods a total of four training set were
composed that include varying illumination, expression
and pose factors for each person and remain factors are
chosen as the test set. The rate of training and testing are:
Case 1: 15% of the dataset for testing and 85% for training.
Case 2:20% of the dataset for testing and 80% for training.
Case 3: 30% of the dataset for testing and 70% for training.
Case 4:40% of the dataset for testing and 60% for training.
According to experiments that included these edge
detection filters, eigenfaces based on PCA technique, SLP
method and ANN we registered the results. In the
following paragraphs, we are showing it:
7.1 Laplacian Based Operators
The entire Laplacian Based Operators algorithms are
designed to depress noise in the first step, and then extract
images edges. This producer is best when we work on
segments of images instead of the image to save fractions
details. Those second order filters can save segments
details more than first order operators. Zero crosses (ZF)
using Laplacian function to discover the edges points.
Where the Laplacian value crosses zero region, it signs
this regain as an edge. Generally, zero crosses know as
general edge detection class that includes log filter and
another. Mathematically,
Fig. 3 Performance of Proposed System based on Zero cross
Laplacian of Gaussian defines the best kernel of edge
detection by measuring the ratio of the signal to noise of
the image. Edge detection by the Laplacian of an image
means taking its double derivative process in both
directions, horizontally and vertically. And because of it is
a member of second-order derivative filters, it has a
stronger response to smooth details. Consequently, well
detection operators when we are using segmentation
process. Table(2) includes classification results based on
LOG, SLP, PCA, and ANN.
Fig. 4 Performance of Proposed System based on LOG, DWT, and SLP
Canny filter (CF)includes three processes. It is performing
noise depressing process firstly by using the Gaussian
function. Then, perform a process to detect the top value
of first derivative operation also corresponds to the
minimum of the second derivative by using threshold
value. The power of canny filter algorithm comes from
this point. This step enables canny algorithm to detect
weak and strong edge points in all directions because it is
not susceptible to noise as compared with other edge
detection operators. Table (2) illustrate the best results
according to experimental tests.
Fig.5 Performance of Proposed System based on Canny, DWT, and SLP
IJCSNS International Journal of Computer Science and Network Security, VOL.18 No.3, March 2018
175
Daubechies wavelets were performed with level (21 -23)
decomposition for getting the optimal choice to create
feature vectors. Thus, we can see the differences between
the algorithms (as it is explained in table 1 2 and 3) and
make a fair comparison among them. Each of dataset
images was converted to four sub-bands LL, LH, HL, and
HH. The LL image represents the main image
(approximate image) and it includes sufficient details that
can be used to reconstruct the original image and ignores
all three subbands, for data reduction in our proposed
algorithm.
Table 1: ZF-Zeroross Filter, DWT-Discrete Wavelet Transformation
SLP-Slope Based Method, ANN
TECHNIQUE TRAINING
DB. (%)
T.C.R.
(Training
Classificat
ion Rate)
T.C.R. (Test
Classificatio
n Rate)
ZF-DWT-
SLP-ANN
0.85 97.6534 96.6667
ZF-DWT-
SLP-ANN
0.8 98.6564 93.5484
ZF-DWT-
SLP-ANN
0.7 98.6784 92.5373
ZF-DWT-
SLP-ANN
0.6 98.9744 92.2680
ZF-DWT-
SLP-ANN
0.5 99.0854 90.6736
Table 2: LF-LOG Filter, DWT-Discrete Wavelet Transformation SLP-
Slope Based Method, ANN
TECHNI
QUE TRAINING
DB. (%)
T.C.R.
(Training
Classification
Rate)
T.C.R. (Test
Classification
Rate)
LF-DWT-
SLP-ANN
0.85 97.6534 96.6667
LF-DWT-
SLP-ANN
0.8 98.6564 93.5484
LF-DWT-
SLP-ANN
0.7 98.6784 92.5373
LF-DWT-
SLP-ANN
0.6 98.9744 92.2680
LF-DWT-
SLP-ANN
0.5 99.0854 90.6736
Table 3: CF-Canny Filter, DWT-Discrete Wavelet Transformation SLP-
Slope Based Method, ANN
TECHNIQUE TRAINING
DB. (%)
T.C.R. (Test
Classificatio
n rate)
T.C.R. (Test
Classificatio
n Rate)
CF-DWT-
SLP-ANN
0.85 97.6534 96.6667
CF-DWT-
SLP-ANN
0.8 98.6564 93.5484
CF-DWT-
SLP-AN
0.7 98.6784 92.5373
CF-DWT-
SLP-ANN
0.6 98.9744 92.2680
CF-DWT-
SLP-ANN
0.5 99.0854 90.6736
PCA Method
The principal components analysis is a very powerful tool
in feature extraction field, especially in the face
recognition. It is wildly used in the face recognition
technology for feature extraction and dimensionality
reduction. We used it for same purposes, directionality
redaction addition to feature extraction with different edge
images. Applied 6 edge detection filters with PCA
increasing the area of study to reach to the optimal pre-
processing and features extractors combination. This
cognition made the training of data set an efficient. BIO ID
database is very complex database because of the
similarity between foreground and background data. The
PCA method that used in the proposed system has shown
good result despite the change in light conditions, face
expression, and complex background of the database
Images.The idea of this method is very simple and
efficient. According to the edge detection filters on the
dataset images, the process of using edge point positions
instead of edge values is the optimal solution to find face
template. This is first and an important operation instead
of using face detection methods that need more time,
efforts, and storage. The output of edge detection filters is
points or scatters with (0,1) value. Consequently, from
these output points, we are visualizing faces by saving
only 1s values. This process is data reduction process
because of excluding background values from all dataset
images. Instead of using these edge points values we used
positions (x,y)of points values. Proposal system created
positions matrix of pixels values to calculate the slope
between every two points and save the results.
8. conclusion
we applied slope method as features vectors. This method
is robust against environmental factors like light condition,
complex background. Also, it doesn’t need complex
normalization process such as dimensionality reduction.
We applied this feature extractors method with same edge
detection filters to measure the efficiency of the system
from PCA, SLP techniques view. This paper found the
importance of the relation between pre-processing
operation, features extractors output vectors that used as
input to the neural network. According to the results of
proposal system in testing level, we registered correct
classification rate practically. The inputs of ANN consist
of features vectors that extracted by SLP and PCA
techniques. A number of neurons that has critical effects in
the hidden layer is an open parameter in the proposed
ANN (it is changeable according to the differences
between actual output and the desired output). The
sigmoid function used in the hidden layer. A range of
neural network cycle(1-10000 epochs )is to measure the
rate of training, testing errors in each experiment. We
tested 3000 experiments in this paper with different meta
parameters like
•
Many edge detection methods.
•
The range of threshold values (0.01-0.9).
IJCSNS International Journal of Computer Science and Network Security, VOL.18 No.3, March 2018
176
•
No of hidden units (25,50,75,100,125,150) .
•
several segments scenario (for example 4 ,6,8,10) .
The training rate of data set to test data set was:
•
85 % from the dataset for training 15% for testing.
•
80 % for training data set 20% for testing.
•
70 % for training data set 30% for testing.
•
60 % for training data set 40% for testing.
As a future work, we are studying to try various discrete
wavelet transformation (DWT) with gradient, Laplacian
filters to improve the data reduction process and for
performance enhancement. Also, a different type of neural
network like convolution neural network can take place in
the new proposed algorithm. Our results indicate that
better recognition rates are obtained with SLP method.
This method can be used for features extraction in another
field as well.
References
[1] Turk, M. a., & Pentland, A. P. (1991). Face
Recognition Using Eigenfaces. Journal of Cognitive
Neuroscience https://doi.org/10.1109/CVPR.1991.139758
[2] Bartlett, M. S. (2001). Independent Component
Representations for Face Recognition. Face Image Analysis
by Unsupervised Learning, 3967. https://doi.org/
10.1007/978-1-4615-1637-8-3
[3] Belhumeur, P. N., Hespanha, J. P., & Krieg man, D. J.
(1997). Eigenfaces vs. fisher face: Recognition using class
specific linear projection. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 19(7), 711720.
https://doi.org/10.1109/34.598228
[4] Lu, J., Member, S., Platinoids, K. N., &
Venetsanopoulos, A. N. (2003). Face Recognition Using
Kernel Direct Discriminant Analysis Algorithms. Analysis,
14(1), 117126. http://doi.org/10.1109/ TNN. 2002.806629
[5] Lawrence, S., Giles, C. L., Ah Chung Tsoi, & Back, A. D.
(1997). Face recognition: a convolutional neural-network
approach. IEEE Transactions on Neural Networks, 8(1),
98113. https://doi.org/ 10.1109/ 72. 554195
[6] Karatzoglou, A., Meyer, D., & Hornik, K. (2006). Support
Vector Algorithm in R. Journal of Statistical Software, 15(9),
128.
[7] DISCRETE WAVELET TRANSFORMS ͳ THEORY AND
APPLICATIONS EDITED BY JUUSO OLKKONEN.
(N.D.).
[8] UrÅun, O., & Favorov, O. V. (2004). SINBAD automation
of scientific discovery: From factor analysis to theory
synthesis. Natural Computing, 3(2), 207233. https://doi.org/
10.1023 /B: NACO .0000027756 .50327.26
[9] Alazzawi, Osman, Bayat, face recognition based on multi
features extractors” IEEE, (2017).
[10] http://vis-www.cs.umass.edu/lfw/
[11] http://www.pitt.edu/ emotion/ck-spread.htm
[12] http://www.jdl.ac.cn/peal/
[13] http://www.kasrl.org/jaffe.html
[14] BioID-Face-Database @ www.bioid.com. (n.d.). Retrieved
from https:// www.bioid.com /About/ BioID-Face-Database
[15] Gonzalez, R., & Woods, R. (2002). Digital image
processing. Prentice Hall. https://doi.org/10.1016 /0734-
189X(90)90171-Q.
[16] Chaple, G. N., Daruwala, R. D., & Gonane, M. S. (2015).
Comparisons of Robert, Prewitt, Sobel operator based edge
detection methods for real time use on FPGA. Technologies
for Sustainable Development (ICTSD), 2015 International
Conference on, (1), 47. https://doi.org/10 .1109/
ICTSD.2015.7095920
[17] Maini, R., & Aggarwal, H. (2009). Study and comparison of
various image edge detection techniques. International
Journal of Image Processing, 3(1), 111. https://doi.org/
http://www. doaj.org /doajfunc-openurl&genre-article&issn-
19852304&date-2009&volume-3&issue-1&spage-1
[18] Gllillon, s., & donias, m. (n.d.). directional second order
derivatives: application to edge and corner detection i :, 2(3),
3–6.
[19] Wei-Bo, Y., & Zun-Sheng, D. (2011). An improved Kirsch
human face image edge-detection method based on a canny
algorithm. 2011 International Conference on Consumer
Electronics, Communications and Networks,
[20] Spontón, H., & Cardelino, J. (2015). A Review of Classic
Edge Detectors. Image Processing on Line, 5, 90–123.
https://doi.org /10 .5201/ipol.2015.35
[21] Revisited, S., & Sorting, P. (2004). Introduction to
Algorithms Part 1 : Divide and Conquer Sorting and
Searching. Sort.
[22] Edelsbrunner, H., Assistant, T., & Gu, Z. (2008). Design
and Analysis.
[23] Barford, L. A., Fazio, R. S., & Smith, D. R. (1992). An
introduction to wavelets. Hewlett-Packard Labs, Bristol, UK,
Tech. Rep. HPL-92-124, 2, 129.
https://doi.org/10.1109/99.388960
[24] Yamamoto, A., & T. L. Lee, D. (1994). Wavelet Analysis:
Theory and Applications. Hewlett-Packard Journal,
(December), 44–52. https://doi.org/ 10. 1051/ jp1:1997114
[25] Discrete, T. H. E., & Transform, W. (n.d.). No Title, 1–15.
[26] Alorf, A. A. (2016). Performance evaluation of the PCA
versus improved PCA (IPCA) in image compression, and in
face detection and recognition. 2016 Future Technologies
Conference (FTC), (December), 537–546.
https://doi.org/10.1109/ FTC .2016 .7821659
[27] Kang, J., Lin, X., & Yang, G. (2015). Research of Multi-
Scale PCA Algorithm for Face Recognition.
[28] Kumar, D. S. D., & Rao, P. V. (2015). Analysis and Design
of Principal Component Analysis and Hidden Markov
Model for Face Recognition. Procedia Materials Science,
10(Cnt 2014), 616–625.
https://doi.org/10.1016/j.mspro.2015.06.014.
[29] Liu, K., & Moon, S. (2016). Robust dual-stage face
recognition method using PCA and high-dimensional-LBP.
2016 IEEE International Conference on Information and
Automation (ICIA), (August), 1828–1831.
https://doi.org/10.1109/ ICInfA.2016 .7832115.
[30] Clarkson, T. G. (1996). Introduction to neural networks.
Neural Network World (Vol. 6).
https://doi.org/10.1016/S0140-6736(95)91746-2
IJCSNS International Journal of Computer Science and Network Security, VOL.18 No.3, March 2018
177
Abdulbasit K.S. Alazzwi Completed
M. Sc. from Iraqi Committee of
Computes and Informatics in computer
science. He is currently pursuing the
Ph.D. degree at the electronic and
computer engineering institute-Altinbas
university. His research interests in the
pattern recognition, image processing,
and machine learning.
Osman N. UCAN received the B.SC.,
M.SC. and Ph. Degrees from Technique
Istanbul University(IT), Turkey 1985,1988
and 1995, in respectively. he has published
more than 250 papers in different fields.
Now he is a Professor, Doctorate
Supervisor and the Dean of the faculty of
engineering, Altinbas university. His main
research interests include biomedical,
image processing and machine learning.
Oguz Bayat received the B.SC.,
degree from Istanbul Technique
University(IT), Turkey in 2000.he has got
M.Sc. from University of Hartford,
Electrical Engineering, 2002. completed the
Ph.D. in the North-eastern University,
Electrical Engineering, 2006.extra, he has
got CP from Massachusetts Institute of
Technology, Management, and Leadership,
2009. he has published more many papers
in different fields. Now he is assisting. Professor, Doctorate
Supervisor and the Dean of the graduate institute, Altinbas
university. His main research interests include signal processing,
communication, and machine learning
