A Survey on Comparison of Face Recognition

Algorithms

Ahmet Özdil Metin Mete Özbilen

Dept. of Computer Engineering Dept. of Computer Engineering
Abdullah Gul University Meliksah University
Kayseri, Turkiye Kayseri, Turkiye
[email protected] [email protected]

Abstract— The identification of a face from a video or image

source is a study of computer vision know as face detection or II. METHODS USED
recognition. Face detection and recognition becomes popular in There are many methods used in face recognition. Each has
recent years by the development of computing power. In this study different features under different conditions like illumination,
we will present performance aspect of algorithms Eigenfaces,
Fisherfaces, and Local Binary Pattern Histograms in different
expression and pose change. Among them, for our purpose
development platforms: Arm and Intel processors. which is face detection in video stream, we focused on three
major methods and we looked their performance under different
Index Terms—Embedded, Face Recognition, Eigenfaces, development environment to find a good matching and low
Fisherface, LBPH. (key words) computing costs.
We tried Eigenfaces, Fisherfaces, and Local Binary Pattern
I. INTRODUCTION Histograms (LBPH) methods to compare with each other. Both
The identification of a face from a video or image source is Eigenfaces and Fisherfaces methods are one of the well-known
a study of computer vision know as face detection or techniques for face recognition. They are known to be very
recognition. Face detection and recognition becomes popular in sensitive to pixel level variations such as illumination, facial
recent years by the development of computing power. It has expression and pose variations. LBP is a relatively new method
wide range of applications including: biometrics, content-based primarily developed for texture analysis. Compared to
image retrieval systems, photography and video processing. Eigenfaces and Fisherfaces methods, we found that LBPH is
Computing and visual sensing technologies in today’s world significantly more robust under illumination and pose
has reach to a state that inexpensive, reliable and accurate variations.
solutions can be feasible. Many embedded systems supplied
A. Eigenfaces
with CMOS camera can be used in face detection systems, and
because of the heavy computing process, a developed versions Eigenfaces is based on Principal Component Analysis
of this systems can be used in face recognition. (PCA). PCA is used to reduce the dimension of an image matrix.
One of major application of face recognition system is For example if face images are represented in g-dimensional
biometric devices and many studies are going on, a different space, PCA uses a linear transform and aims to get an h-
type of application can be used in segmenting a movie among dimensional subspace, which answers maximum variance in the
with its cast. This application can be used in jumping any scene g-dimensional space and where h is too small according to g.
of your favorite actor/actress. Mean centered images are calculated by subtracting the
Television or media player are equip with enhanced normalized training images from the calculated mean image. If
embedded systems, which can perform face recognition off-line W is the matrix of mean centered training images Wi (i = 1, 2, ...
or on-line. L) and L is the number of training images, covariance matrix D
When the job is recognition of a face in a movie, the process is calculated from W as in Equation 1.
becomes more complicated, since the images you have to
process are generally not in an ideal condition like in biometric  =  
devices. The person you have to detect and recognize can be in
various conditions. These conditions can be low illumination, To reduce the size of covariance matrix D, we can use D =
contrast and scene conditions. WTW instead. Eigenvectors ei and eigenvalues λi are obtained
This study we will present performance aspect of different from covariance matrix.
algorithms in different development platforms.
 ( = 1,2, … , ) (2)

In the Equation 2, zi represents the new feature vector of the

new lower dimensional subspace [1].
There is a negative aspect of this method, that it tries to
maximize inter and intra class scattering. Inter class scattering is
good for classification while intra class scattering is not. In face
recognition, if there is variance of illumination, this increases
intra class scattering very high, even classes seems stained, and
causes low classification[2]. Fig. 2. ELBP operator examples [5]

B. Fisherfaces The histograms of LBP are used for face recognition since
Linear Discriminant Analysis (LDA) or Fisherfaces aims to LBP histograms contain information about the distribution of
increase inter class differences, not data representation. local micro patterns. Because the face image is too big for LBP
calculation, dividing the image into small regions is proposed in
[6]. Some parts of face (like eyes, mouth) contain more
= ∑ ∑ ( − )( − ) (3)
information for face recognition. Yang et al. proposes to train
and allocate different weights for face parts, by their information
 = ∑ ( − )( − ) (4) covering and then concatenating them end to end to build up
global description of face. This helps to collect local pattern
The formulas above are intra class (Equation 3) and inter information with spatial details of the whole image.
class (Equation 4) scatter matrices respectively. The indices, j is To decide if two face images are belong to same person, the
class and i is image number. μj is the mean of class j, and μ is images' histograms are compared. Chi square statistic similarity
mean of all classes. Mj shows the number images in class j, and measure is used for comparison of histograms. It can be defined
R is the number of classes. Sb is maximized while Sw is as follows:
minimized for the classification to be done [1].
2
C. Local Binary Patterns (LBP) 2
( , ) = ∑,
( ,− , )
(5)
( ,− , )
Local Binary Patterns (LBP) was first presented by Ojala et
al. in [3] to use in texture description. The basic method, labels
each pixel with decimal values called LBPs or LBP codes, to where i = 0,1,...,n-1, j = 0,1,...,m-1, wj is the weight for region j,
describe the local structure around of pixel. As illustrated in S is target face image histogram and M is the query face image
Figure 1, value of the center pixel is subtracted from the 8 histogram[6].
neighbor pixels' values, if the result is negative the binary value III. EXPERIMENTS
is 0, otherwise 1. The calculation starts from the pixel at the top
left corner of the 8-neighborhood and continues in clockwise In our study two experiment setups are used. The faces of the
direction. After calculating with all neighbors, an eight digit TV series Buffy season 5 episode 2 are extracted and three main
binary value is produced. When this binary value is converted to characters' faces are used in the experiments. Buffy, Michelle
decimal, the LBP code of the pixel is generated, and placed to and Nicholas are the characters that are chosen. Nicholas is
the coordinates of pixel in matrix [4]. chosen to add at least one male character to testbed. As for male
faces are better suited for face recognition algorithms, and thus
give better results.
The aim of the study is comparing the face recognition
performances of the selected the three methods. So two different
testbed is prepared accordingly.

TABLE I. EXPERIMENT TESTBEDS

Fig. 1. Basic LBP operator [4]
Experiment 1 Experiment 2

There is a drawback of LBP which uses 8-neighborhood Characters Train(# of Test(# of Train(# of Test(# of
face face face face
(3x3) that cannot cover large-scale structures. To take into images) images) images) images)
account texture of different size structures, the method is
generalized. In [5] Ojala et al. revised the method to be flexible Buffy 14 739 451 302
for any radius and any number of sampling points and named Michelle 18 606 375 249
the new method as Extended LBP(ELBP). Figure 2 shows
different examples of ELBP operator, and P represents the Nicholas 13 192 123 82
number of neighbors and R represents the radius of a circle on
which neighbors are located. The difference between the experiments is that, in
experiment 1, one track of a face is chosen as training set and
the rest as test set. In experiment 2, 60% of the images is chosen In experiment 1, one face track is used, this causes a decrease
as training set and 40% as test set. in the success of recognition dramatically to the level of random
OpenCV [7] library is used for image processing within the selection. Selecting only one track causes the train set to be
experiments. Microsoft Visual C++ 2010 Express Edition is inefficient. But in experiment 2, 60% of the whole images are
used for coding and compiling on windows based system. And used, and the training set is selected from all tracks.
for arm-Linux system g++ is used to compile the code.
VI. CONCLUSION
IV. PLATFORMS In the article, Eigenfaces, Fisherfaces, and Local Binary
We run same algorithms on different platforms, to show how Pattern Histograms (LBPH) methods in face recognition is
the algorithms benefit from the hardware and see performance compared with each other in different hardware environments.
variance. The result will guide anyone, who wants to build an Their recognition success and performance are measured.
embedded face recognition system. The results showed that Fisherface gives best timing while
Two different platforms are used. First platform is Intel lack of successful recognition and LBPH has succeed in
based running at 2.7 GHz processor with 4 GB ram, labeled as recognition with the expense of performance.
Intel. Second platform is arm based running at 1 GHz processor
512 MB ram, labeled as Arm. ACKNOWLEDGMENT
This project is granted by The Ministry of Science, Industry
V. RESULTS and Technology Turkiye, with grant number 0370.STZ.2013-2.
The timing performance of the algorithms are shown in
Table 2. REFERENCES
[1] M. Sharkas, M. Abou Elenien, “Eigenfaces vs. Fisherfaces vs.
TABLE II. EXPERIMENT TIME DURATIONS ICA for Face Recognition; A Comparative Study,” 9th
International Conference on Signal Processing, 2008, ICSP
Methods Experiment 1 Experiment 2 2008., 2008, pp. 914–919
Intel Arm Intel Arm [2] P. N. Belhumeour, J. P. Hespanha, and D. J. Kriegman,
(seconds) (seconds) (seconds) (seconds) “Eigenface vs. Fisherfaces: Recognition Using Class Specific
Eigenfaces 1.91 30.65 432.08 3600 Linear Projection”, IEEE Trans. on PAMI, 19(7), pp. 71 I-720,
1997
Fisherfaces 5.96 8.3 549.84 3360 [3] T. Ojala, M. Pietikäinen, and D. Harwood (1994), "Performance
LBPH evaluation of texture measures with classification based on
55.02 136.52 143.61 907.02
Kullback discrimination of distributions", Proceedings of the 12th
IAPR International Conference on Pattern Recognition (ICPR
As seen in Tab 2, in experiment 1, Eigenfaces made the best 1994), vol. 1, pp. 582 - 585
score of all three, especially by far to LBPH. But Fisherfaces [4] D. Huang , C. Shan , M. Ardabilian , Y. Wang and L. Chen
also gave good result with 5.96 seconds. But overall LBPH is "Local binary patterns and its application to facial image analysis:
robust to change in the hardware architecture. Especially in Arm A survey", IEEE Trans. Syst., Man, Cybern., C, vol. 41, pp.765
platform, LBPH is made the best results. -781 2011
[5] T. Ojala, M. Pietikäinen, and T. Maenpaa, “Multiresolution gray-
TABLE III. EXPERIMENT TRUE/FALSE ANSWERS scale and rotation invariant texture classification with local binary
patterns,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no. 7,
Methods Experiment 1 Experiment 2 pp. 971–987, Jul. 2002.
True False Hit Ratio True False Hit Ratio [6] H. Yang and Y. Wang, “A LBP-based face recognition method
(%) (%) with Hamming distance constraint,” in Proc. Int. Conf. Image
Eigenfaces Graph., Aug., 2007, pp. 645–649.
509 1028 33 586 47 93
[7] The OpenCV Library Dr. Dobb’s Journal of Software Tools, G.
Fisherfaces 519 1018 34 540 93 85 Bradski, 2000.
LBPH 679 858 44 594 39 94