VOL. 9, NO.

10, OCTOBER 2014 ISSN 1819-6608

ARPN Journal of Engineering and Applied Sciences
©2006-2014 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

FACE RECOGNITION USING SCAN-BASED LOCAL FACE DESCRIPTOR

Riko Arlando Saragih, Dodi Sudiana and Dadang Gunawan
Department of Electrical Engineering, University of Indonesia, Indonesia
E-Mail: [email protected]

ABSTRACT
This paper describes SCAN descriptor as a local face descriptor to represent a face image. SCAN techniques that
originally for image compression and data hiding were used to locally extract face image features to represent the face
image. Simulations were conducted on the subset of cropped Yale Face Database B by either varying uniformly the face
image pixels (intensities) or lowering their resolutions in the database subset. The simulation results show that SCAN
descriptor has recognition rate that outperforms for both either two global face descriptors, i.e. Principal Component
Analysis (PCA) and Linear Discriminant Analysis (LDA), or two local face descriptors, i.e. Local Binary Pattern (LBP)
and Multi-Scale Local Binary Pattern (MLBP).

Keywords: SCAN, local face descriptor, global face descriptor, recognition rate.

INTRODUCTION resolution, pose, and occlusions, are among factors that

The rise of terrorist acts as criminal conduct or deteriorate recognition rate. This situation exacerbated
acts of bombing protesters / motorcycle gang that often act generally only a single face image that available for
anarchist or theft at a supermarket / bank / parking area training phase [3, 5].
requires two approaches, namely prevention through a Face descriptor can be divided into two
more rigorous examination of persons suspected of categories, namely global face descriptor and local face
(preventive) and search for people who are suspected of descriptor. Global face descriptor is obtained by
having committed a crime (curative). To facilitate this processing the entire face (integral/whole face image)
search, one needs a system that can assess the similarity directly to obtain important information from a face. Two
among subjects that are monitored against the existing famous global face descriptors are PCA and LDA [6, 7].
database. In contrast, local face descriptor typically
Face, as one of biometrics, is a prominent and obtained by feature extraction from specific
easiest information that needed this matching process. components/regions of a face. In the development, along
This is because face is a form of personal identification with the need for a face recognition system that is real
based on biometric person who found in him, not what he (typical practical), then the local facial image descriptors
owns (e.g. ID card) or what he remembers (password) [1]. received attention from researchers and attempt to develop
Therefore, as a human being, anywhere, anytime, under a local facial descriptor that is more robust against to
any conditions, someone automatically always bring their variations in lighting, occlusions, and the blurred facial
identity information in the form of face. images (low-resolution)[8]. LBP and its successor, MLBP,
Face recognition accuracy is determined by a can be named as the famous corner stone among local face
description / representation of face images and the design descriptors [9, 10].
of classifiers. The purpose of representation is to obtain a
set of facial feature (characteristic) of a face image which SCAN METHODOLOGY
minimizes the intra-class variations (e.g. between different A scanning of a two dimensional array Pmxn =
facial image but from the same individual), while {p(i,j) : 1  i  m, 1  j  n } can be considered as a
maximizing the inter-class variation (e.g. between image bijective function that maps every element of Pmxn into a
different individual faces). Meanwhile, essentially a set of distinctive one dimensional array Q = {1,2, ..., mn -
classifier is a function that discriminates a set of novel 1, mn}[11]. In other words, a scanning of a two
facial features to determine its identity. It is important to dimensional array is an order in which each element of the
note that in representing a face, if the representation of the array is accessed exactly once.
face is not robust enough, even the most sophisticated Basically, SCAN is a family of formal languages-
classifiers will fail to perform the role for face recognition. based two-dimensional spatial as a generic methodology
Therefore, it is important to carefully define the for accessing a large number of of wide variety of
representation of the face that will be taken when scanning paths easily [11]. It has several versions, such as
designing a face recognition system [2]. Simple SCAN, Extended SCAN, and Generalized SCAN,
Generally, face recognition problem can be each of which can represent and generate a specific set of
formulated as : given static (still) or video images of a scanning paths. It also has a set of basic scan patterns, a
scene as a novel pattern, identify or verify it by comparing set of transformations, and a set of rules to compose
with patterns stored in a database [3]. As mentioned in [4], simple scan patterns to obtain complex scan patterns [11].
typical practical face recognition problem is a set of These basic scan patterns are shown in Figure-1.
unconstrained conditions, where lighting, image

1779
 VOL. 9, NO. 10, OCTOBER 2014 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2014 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

There are four steps for face image representation

by using SCAN descriptor. The first three steps works on
each block of a face image. These steps can be
summarized as follows.
First, for each scan path, do scanning and
calculate the cumulative absolute difference according to
the scanning path that is working. Second, choose the scan
path that produces the minimum cumulative absolute
difference as the best scan path. Third, encode the best
scan path as a feature using binary code. This binary code
represents each block of a face image. Table-2 shows the
pair of each scan path and its binary code. Ultimately, to
Figure-1. Basic SCAN patterns [11]. represent the whole face image, concatenate all the best
scan path codes as a sequence of binary code feature
SCAN-BASED LOCAL FACE DESCRIPTOR vector. These steps were run both for training phase to
This paper uses twelve scanning paths, each extract information from each image in the database and
working along whole non-overlapping face image blocks, for testing (classifying) phase, where finally a novel
where each block has 4 by 4 pixels. The reason for sequence of binary code was compared with the stored
choosing these SCAN patterns is due to they generate best scan path codes and the face recognition system will
robust discriminative local-features which is characterized identify it as one in the database that has the maximum
by the resulting minimum cumulative absolute difference number of binary code that matches for each
(error). Table-1 depicts all scanning paths that used to corresponding block face image.
extract features from each face image.
Table-2. Binary code for each scanning path.
Table-1. The twelve scanning paths. No Name of Scanning Path Binary Code
No Type and Name of No Type and Name of 1 SCANd0 0000
Scanning Path Scanning Path
1 7 2 SCANd1 0001
3 SCANd2 0010
4 SCANd3 0011

SCANd0 SCANd6 5 SCANd4 0100

2 8 6 SCANd5 0101
7 SCANd6 0110

SCANd1 SCANd7 8 SCANd7 0111

3 9 9 SCANd8 1000
10 SCANd9 1001
11 SCANd10 1010
SCANd2 SCANd8 12 SCANd11 1011
4 10

We illustrate this error pattern coding scheme as

follows. Let I(x,y) is a face image with dimension mxn,
where m (m = 2k) and n (n = 2l) are the width and the
SCANd3 SCANd9
height of the face image, respectively, and k does not need
5 11
to be equal with l (k, l = 2, 3, 4,...). To extract the local
features, first, we divide the face image into several non-
overlapping blocks, where each block has 4 by 4 pixels,
SCANd4 SCANd10 which results in L blocks (L = (mxn)/16). Subsequently,
6 12 we evaluate each block by using all scan paths. Figure-
2(a), Figure-2(b), and Figure-2(c) depict the block face
image before scanned, scan path type that is working (e.g.
SCANd0), and the block face image after scanned,
SCANd5 SCANd11 respectively.

1780
 VOL. 9, NO. 10, OCTOBER 2014 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2014 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

simulation. Figure-4 and Figure-5 display the sample face

images for these, respectively.

Figure-2. Block face image before and after scanned.

Figure-4. Samples of the intensity-reduced face image.
For j = 1, 2,..., 12, to calculate the cumulative
absolute difference for the resulting face image block after
scanned can be expressed as:

15
e j   pk' 1  pk' (1)
k 1

Eventually, the scanning path j that resulting the Figure-5. Samples of the lower-resolution face image.
minimum cumulative absolute difference (ej) was decoded
as listed in Table-2. SIMULATION RESULTS AND DISCUSSIONS
Table-3 and Table-4 show the results for the first
SIMULATION PROCEDURES and second simulation, respectively. Instead of using
Yale Face Database B is one of standard database recall-precision as in [8], the recognition rate was
to evaluate any face recognition algorithm (method), calculated by dividing the number of true match-identity
especially for studying illumination effects on face with the number of face images in database subset that
recognition [12]. Subset of this database that we used for used for each simulation, i.e. 4560 (2x38x60) for the first
simulations contains cropped grayscale face images that one and 2280 (38x60) for the second one.
consists of 38 subjects, each has 60 face images. Each
cropped face image has dimension 168 by 192 pixels. For Table-3. The first simulation results.
every subject, we consider only one pose (frontal) among
nine poses. We pick only one face image for each subject Testing
Descriptor Descriptor Recognition
in the training phase to get its face descriptor. Figure-3 time
type name rate (%)
shows the samples of cropped face images used to (seconds)
evaluate our method. Global PCA 4.74 9.95
Descriptor LDA 17.48 10.14
LBP 12.37 17.50
Local MLBP 45.33 4397.08
Descriptor SCAN 59.56 325.56

Table-4. The second simulation results.

Figure-3. Samples of cropped face images [13].
Testing
Descriptor Descriptor Recognition
We evaluate the descriptor performance time
type name rate (%)
(recognition rate and time needed for testing), either for (seconds)
global (PCA and LDA) or local (LBP, MLBP and SCAN) Global PCA 20.44 3.75
descriptor. The first simulation was conducted to evaluate Descriptor LDA 35.39 3.49
the descriptor performance against pixel intensity changes.
LBP 3.68 6.42
For the first simulation, in the testing phase, we duplicate Local
all the face images in the database subset and also reduce MLBP 44.08 1210.32
Descriptor
uniformly all their pixel intensities with different intensity SCAN 44.34 163.29
scales for each face image.
In the second simulation, we evaluate the For both simulations, it is obvious that every
performance descriptor against blurred face images. In global face descriptor has testing time that faster than
contrast with the first one, we only reduce the resolution of every local face descriptor. This is plausible because as
each face image with scale 0.2 (the resulting face image is described earlier that a global face descriptor works
only one-fifth the original one) in the database subset
without duplicating all the face images for the second

1781
Iteration

VOL. 9, NO. 10, OCTOBER 2014 ISSN 1819-6608

ARPN Journal of Engineering and Applied Sciences
©2006-2014 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

directly on the whole face image rather than local face [2] Stan Z. Li and Anil K. Jain. 2011. Handbook of Face
descriptor that works on block-by-block for a face image. Recognition. 2nd Ed. Springer-Verlag London Limited.
As stated earlier, the first simulation was pp. 79-80 .
conducted by reducing uniformly all pixel intensities with
different intensity scales for each face image in the testing [3] A.S. Tolba, A.H. El-Baz and A.A. El-Harby. 2006.
phase. Due to all the face descriptors are appearance-based Face Recognition: A Literature Review. International
[6, 7, 9, 10], it is obvious that in general the local face Journal of Signal Processing. 2(2): 88-103.
descriptor is better than the global descriptor. It means that
in general the local descriptor may adapt the pixel [4] G. Hua, M.-H. Yang, E. Learned-Miller, Y. Ma, M.
intensity changes rather the global descriptor. Among Turk, D.J. Kriegman and T. S. Huang. 2011.
local descriptors, SCAN descriptor as a local face Introduction to the Special Section on Real-World
descriptor is more robust against pixel intensity changes Face Recognition. IEEE Transactions on Pattern
compared to the other descriptors. Analysis and Machine Intelligence. 33(10): 1921-
The second simulation was conducted to evaluate 1924.
each face descriptor against blurred face images. It might [5] Xiaoyang Tan, Songcan Chen, Zhi-Hua Zhou and
happen whether the camera may be out of focus or the Fuyan Zhang. 2006. Face Recognition from a Single
distance between the camera and the subject being Image per Person: A Survey. Pattern Recognition.
observed is not close enough. It seems not all local face 39(9): 1725-1745.
descriptor have better performance in term of recognition
rate than global face descriptor. As described in [9], LBP [6] M. Turk and A. Pentland. 1991. Eigenfaces for
as a local face descriptor has good performance only for recognition. Journal of Cognitivie Science. 3(1): 71-
monotonic intensity changes, not for a noisy (blurred) one. 86.
It is also suprisingly that both PCA and LDA get higher
recognition rate than the first simulation. Their methods [7] P. N. Belhumeur, J.P. Hespanha and D. J. Kriegman.
that based on preserving the most significant eigenvalues 1997. Eigenfaces vs. Fisherfaces: Recognition using
might give a more discriminating power that leads in class specific linear projection. IEEE Transactions on
better recognition rate. But overall, as in first simulation, Pattern Analysis and Machine Intelligence. 19(7):
SCAN descriptor as a local face descriptor is more robust 711-720.
against low resolution (blurred) face image compared to
the other descriptors. [8] Krystian Mikolajczyk and Cordelia Schmid. 2005. A
For both simulations, we use four radii for the performance evaluation of local descriptors. IEEE
local decriptor MLBP. We choose these to accomodate the Transactions on Pattern Analysis and Machine
multi-scale of the face image in order to gain more Intelligence. 27(10): 1615-1630.
discriminative power. Although in the second simulation [9] T. Ahonen, A.Hadid and M. Pietikainen. 2006. Face
its performance is rather similar with SCAN descriptor, but description with local binary patterns: Application to
MLBP is more time consuming than SCAN. face recognition. IEEE Transactions on Pattern
Analysis and Machine Intelligence. 28 (12) : 2037-
CONCLUSIONS 2041.
We have conducted two kinds of simulations for
each face descriptor, either global (PCA and LDA) or [10] C.H. Chan, J.V. Kittler and K. Messer. 2007. Multi-
local (LBP, MLBP and SCAN). From both simulations, it scale local binary pattern histograms for face
is obvious that SCAN as a local face descriptor has the best recognition. In: Proc. International Conference on
performance in term of recognition rate among face Biometrics. pp. 809-818.
descriptors.
[11] S.S. Maniccam and N. Bourbakis. 2004. Image and
ACKNOWLEDGEMENT video encryption using SCAN patterns. Pattern
The first author is on leave from Maranatha Recognition. 37(4): 725-737.
Christian University and expresses his sincere thanks to
[12] Kuang-Chih Lee, Jeffrey Ho and David Kriegman.
the Indonesian Ministry of National Education for
2005. Acquiring Linear Subspaces for Face
supporting his doctoral study in the form of Higher
Recognition under Variable Lighting. IEEE
Education Scholarhsip (BPPS).
Transactions on Pattern Analysis and Machine
Intelligence. 27(5): 684 - 698.
REFERENCES
[13] Athinodoros S. Georghiades, Peter N. Belhumeur and
[1] Anil K. Jain, Arun Ross and Salil Prabhakar. 2004. An David J. Kriegman. 2001. From Few to Many:
Introduction to Biometric Recognition. IEEE Illumination Cone Models for Face Recognition under
Transactions on Circuits and Systems for Video Variable Lighting and Pose. IEEE Transactions on
Technology. 14(1): 4-20. Pattern Analysis and Machine Intelligence. 23(6):
643-660.

1782