A Survey on Local Feature Based Face Recognition Methods

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 03 Issue: 01 | Jan-2016 www.irjet.net p-ISSN: 2395-0072
© 2016, IRJET | Impact Factor value: 4.45 | ISO 9001:2008 Certified Journal | Page 706
A SURVEY ON LOCAL FEATURE BASED FACE RECOGNITION METHODS
Neethu Krishnan s1, Sonia George2
1 Student, Computer science and Engineering, Lourdes matha college of science and technology, Kerala,India
2 Assistant professor, Computer science and Engineering, Lourdes matha college of science and technology, Kerala
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - – Great progress has been achieved in face
recognition. But it is still challenging to differentiate
the people with similar appearance and to recognize
the same person who has different appearance due to
different facial expression, pose, occlusion, illumination
and aging. This paper provides some of the local
feature based methods that tackle these problems. The
two main approaches of local feature analysis are LBP
and Gabor wavelets. The extension of the LBP is LTP.
The combination of both LBP and Gabor wavelets
provide more discriminative information which
improves the face recognition performance. Combining
information from different domain is beneficial for FR.
Two methods under this are GV LBP TOP and Effective
GV LBP.Then MBC is provided which has much less time
and space complexity than the widely used Gabor
transformation methods. Then GOM, a promising
solution to handle intra-class variation and inter-class
similarity is provided. Extensive experiments on face
image databases such as FERET demonstrate that GOM
provides better recognition rate as compared to other
methods.
Key Words: Key Words: Local feature analysis, Local
BinaryPattern(LBP),LocalTernaryPattern(LTP),Mono
genic Binary Coding(MBC) ,Gabor Ordinal
Measure(GOM).
1. INTRODUCTION
In today’s networked world, the way of crime is becoming
easier than before. For this reason network security is
becoming a major concern in various fields in the society.
We all know that bank and computer system uses PIN,
password, ID,key for identification and security
clearances. When credit and ATM cards are lost or stolen,
an unauthorized user can use the card with easily guessed
PIN and passwords. Recent cases of identity theft have
heightened the need for technology to prove authenticity.
One such technology that we use nowadays is Biometrics.
Biometric systems identify an individual by both
behavioral and biological characteristics.
Facial Recognition Technology (FRT) has emerged as an
attractive solution to address the identification and
verification of identity claims. The face recognition have
application in various field such as information security,
smart cards, access and border control in e-government, e-
health and e-commerce service, criminal identification,
law enforcement and surveillance purposes such as search
for wanted criminals and suspected terrorist.
Face recognition is chosen over other biometric such as
iris, fingerprint due to number of reasons.1) It requires no
interaction on behalf of the user. 2) It is accurate and
allows for high enrolment and verification rates.3) It does
not require an expert to interpret the comparison result.4)
It is the only biometric that allows to perform passive
identification.5) It is easy to use, embed and also
inexpensive compared to iris and fingerprint.
Face recognition techniques have always been a very
challenging task for researches because of its difficulties
and limitations. One problem of face recognition is the fact
that different faces could seem very similar (inter class
similarity). Therefore a discrimination task is needed to
recognize people with similar appearance. On the other
hand, when we analyze the same face, many
characteristics may have changed (intra person
differences) due to changes in illumination, facial
expressions, occlusion, pose, aging and the presence of
accessories. Therefore how to extract robust and
discriminate features which makes the inter-person faces
compact and enlarge the margin among different persons
become a critical and difficult problem in face recognition.
Feature extraction is one of the most important steps in
the process of face recognition in order to overcome these
problems.
The main feature extraction approaches are subspace
holistic features and local appearance features. Holistic
subspace analysis represents the global appearance of a
human face with its projections on the subspace. A large
and representative training set is needed for subspace
learning. This method performs well under controlled
condition and achieves high recognition rate. But
misalignment of face images may cause significant
degradation of recognition performance. Typical holistic
features include principal component analysis (PCA) [1],
linear discriminate analysis (LDA) [2], independent
component analysis (ICA) [3]etc.

In this paper experiment on FERET database are
considered to analyse the recognition performance of
various methods against different factors such as facial
expression, illumination changes and aging.
fa set, used as a gallery set, contains frontal images of
1196 people.
fb set (1195 images). The subjects were asked for an
alternative facial expression than in fa photograph.
fc set (194 images). The photos were taken under
different lighting conditions.
dup I set (722 images). The photos were taken later in
time.
dup II set (234 images). This is a subset of the dup I set
containing those images that were taken at least a year
after the corresponding gallery image.
2. LOCAL FEATURE BASED METHODS
Using the statistical Local Feature Analysis (LFA)
technique, a set of feature points is extracted from each
face image.Local feature analysis has achieved better
performance in different face recognition tasks which is
characterized by the following reasons: 1) local features
can characterize the facial details for personal
identification; 2) there are some local features designed
with illumination robust property such as LBP.3)
statistical histograms of local features are used as facial
descriptors, being robust against partial variations of local
features; 4) local methods provide more flexibility for
recognizing faces with partial occlusions.
Two representative methods of local feature analysis in
face biometrics are Gabor wavelets [4] and Local Binary
Patterns (LBP) [5]. Gabor wavelets can extract the local
features of facial regions on multiple channels of
frequencies and orientations. Both magnitude and phase
information of Gabor wavelets can be used for face
recognition. Gabor features are discriminative and robust
to illumination and expression changes and it can explore
the neighbouring relationship in spatial, frequency and
orientation domain. Gabor transformation has high
computational cost and storage space because Gabor
transformation of an input image needs to be
implemented in multiple scale and orientation. The
general shortcoming of the Gabor wavelet is high
dimensionality of the feature vector. To address this
problem, dimensionality reduction techniques such as
kernel PCA, LDA etc are used.
2.1. Local Binary Pattern
Local binary pattern is a means of summarizing local grey
level structure. The LBP operator takes a local
neighborhood around each pixel, thresholds the pixels of
the neighborhood at the value of the central pixel and uses
Fig. 1. Illustration of the sensitivity of local binary patterns to noise.
the resulting binary-valued image patch as a local image
descriptor. Initially 3 ×3 neighborhoods are used which
gives 8-bit integer LBP code. Formally, the LBP operator
takes the form,
where in this case n runs over the 8 neighbors of the
central pixel c, ic and in are the gray-level values at c and
n, and s(u) is 1 if u≥0 and 0 otherwise. The LBP encoding
process is illustrated in Fig. 1. LBP features are completely
invariant to monotonic grey scale transformation.
Therefore LBP is simple and effective way to represent
face image because of its robustness with respect to facial
expression, aging, illumination and alignment.
LBP based methods achieved a recognition rate of 93% in
the case of recognizing faces under different facial
expressions (fb set). Under different lighting conditions (fc
set), the LBP has also achieved the best performance with
a recognition rate of 51%. The relatively poor results on
the fc set confirm that illumination change is still a
challenge to face recognition.
2.2. Local Ternary Pattern
LBP have proven to be highly discriminative features for
texture classification and they are resistant to lighting
effects. However because they threshold exactly at the
value of the central pixel ic they tend to be sensitive to
noise, particularly in near-uniform image regions such as
cheeks and foreheads .So a generalization of LBP called
local ternary patterns (LTP)[6] is introduced that is more
discriminant and less sensitive to noise in uniform regions.
LTP extends LBP to 3-valued codes in which gray-levels in
a zone of width ±t around ic are quantized to zero, ones
above this are quantized to +1 and ones below it to -1,
i.e., the indicator s(u) is replaced with a 3-valued function,

Fig 2. Illustration of the basic LTP operator
The binary LBP code is replaced by a ternary LTP
code.Eventhough LTP codes are more resistant to noise,
they are no longer strictly invariant to gray-level
transformations. The LTP encoding procedure is
illustrated in Fig. 2.
2.3 LBP AND GABOR WAVELET
LBP is good at coding fine details of facial appearance and
texture, whereas Gabor features encode facial shape and
appearance over a range of coarser scales. Both
representations are rich in information and
computationally efficient, and their complementary nature
makes them good candidates for fusion. Both Gabor and
LBP feature increases the recognition rate by 30% relative
to LBP features alone and by 10% relative to Gabor
features alone. So the combinations of LBP and Gabor
features have improved the FR performance significantly
compared to the individual representation.
Combining information from different domains is usually
beneficial for face recognition [7].First, the multiscale and
multiorientation representations are derived by
convolving the face image with a Gabor filter bank and
formulated as a third-order volume. Second, LBP operator
is applied on the three orthogonal planes of Gabor volume,
respectively, named GV-LBP-TOP. In this way, neighboring
information is encoded not only in image space but also
among different scales and orientations of Gabor faces.
Advantages
1) Gabor feature is applied to the face images to alleviate
the variations of facial expression and illumination.
2) LBP is utilized to model the neighboring relationship
jointly in spatial, frequency and orientation domains. In
this way, discriminant and robust information, as much as
possible, could be explored.
2.3.1. Gabor Volume Based LBP on Three
Orthogonal Planes (GV-LBP-TOP)
Discriminative information is explored by modeling the
neighboring relationship among spatial, frequency and
orientation domain. For a face image, the derived Gabor
faces are assembled by the order of different scales and
orientations to form a third-order volume as illustrated in
Fig. 3 where the three axes X, Y, T denote the different
rows, columns of face image and different types of Gabor
filters respectively.
Fig.3. Face image and its corresponding third-order Gabor
volume.
It first applies LBP analysis on the three orthogonal planes
(XY, XT, and YT) of Gabor face volume and then combines
the description codes together to represent faces. The
codes from three planes are different and, hence, may
supply complementary information helpful for FR. After
that, three histograms corresponding to GV-LBP-XY,GV-
LBP-XT, and GV-LBP-YT codes are computed.
The GV-LBP-TOP histogram H is finally derived by
concatenating these three histograms H=[H1,H2,H3] to
represent the face.
2.3.2. Effective GV-LBP
The aforementioned GV-LBP-TOP is of high computational
complexity. The length of the histogram feature vector and
the computational cost are threefold compared to those of
LGBPHS. So it is not very efficient in practical application.
To address this problem, we propose an effective
formulation of GV-LBP (E-GV-LBP) which encodes the
information in spatial, frequency and
Orientation domains simultaneously and reduces the
computational cost. Fig. 4 shows the definition of E-GV-
LBP coding. For the central point, Ic, I0 and I4 are the
orientation neighboring pixels; I2 and I6 are the scale
neighboring ones;I1,I3,I5 and I7 are the neighboring pixels
in spatial domains.
Fig. 4. Formulation of E-GV-LBP.

The E-GV-LBP value have the form,
Where S(IP-IC) is a threshold function defined as
Regarding to the complexity and accuracy, E-GV-LBP is the
best method for face representation and recognition.
Unlike PCA, LDA etc. whose performances greatly depend
on the training set, for GV-LBPTOP and E-GV-LBP
methods, once the weight for every block is fixed, no
further training procedure is needed.
2.4. MONOGENIC BINARY CODING
Local feature based face recognition methods, such as
Gabor features encoded by local binary pattern, could
achieve state-of-the-art FR results in large-scale face
database such as FERET. Gabor filter based FR methods
are very expensive: both the computational cost and the
storage space are high because Gabor filtering is usually
applied at five different scales and eight different
orientations.So an efficient local feature extraction
scheme, namely monogenic binary coding (MBC)[8], for
face representation and recognition is proposed.
Monogenic signal representation decomposes an original
signal into three complementary components: local
amplitude, local orientation and local phase. . The local
orientation and local phase compose monogenic phase. It
does not use steerable filters to extract multiple-
orientation features, and thus it has much lower time and
space complexity than Gabor transformations.It encode
the monogenic variation in each local region and
monogenic feature in each pixel, and then calculate the
statistical features (e.g., histogram) of the extracted local
features.To exploit the discrimination information
embedded in the amplitude, phase and orientation
components monogenic binary coding (MBC) is used to
encodes the local pattern in different monogenic feature
maps.
The proposed algorithm of coding the monogenic signal
features contains two parts: monogenic local variation
coding and monogenic local intensity coding. The first part
encodes the variation between the central pixel and its
surrounding pixels in a local patch, while the second part
encodes the value of central pixel itself. Obviously, these
two parts are complementary.
The local amplitude of monogenic signal representation is
a measurement of local energetic information. For
example, high amplitude usually indicates higher energetic
local features (e.g., edges, lines, textures). Therefore, the
local variation of monogenic amplitude could be coded by
comparing the amplitude value of central location with
those of its neighbours.
For 2D signals, monogenic phase is decomposed into local
orientation and local phase. The local orientation
describes the geometric information, more specifically the
main orientation of the local structure, and the local phase
describes the information of the local structure along the
main orientation. The imagery part of monogenic signal
representation is used to encode the local feature intensity
information of the central pixel.
2.4.1. Monogenic binary code (MBC)
For each of the amplitude, phase and orientation
components of the monogenic signal representation, we
can have a monogenic binary code (MBC) map. The code
maps for amplitude, phase and Orientation are denoted by
MBC-A, MBC-P and MBC-O respectively. At each location,
MBC-A, MBC-P and MBC-O are formed by combining the
local imagery intensity code and the local variation code.
After computing the multi-scale MBC-A, MBC-P and MBC-O
feature maps of the face image, we can construct three
histograms, HMBC-A, HMBC-P and HMBC-O, through the
following procedures. For each kind of pattern map on
each scale, it is partitioned into multiple non-overlapping
regions, and then the local histogram is built for each sub-
region. Finally, all the local histograms across different
scales and different regions are concatenated into a single
histogram vector to represent the face image.
The local statistical features extracted from the
complementary monogenic components (i.e., amplitude,
orientation and phase) are then fused for effective FR.
MBC has higher accuracy and low cost than the Gabor
feature methods. Monogenic signal representation can
effectively extract the discrimination information
embedded in the original signal by decoupling the local
energy (local amplitude) and structure (local phase and
local orientation). The monogenic signal analysis is very
suitable for face image representation because human
facial local structures can be well characterized along one
dominant orientation. The experimental results show that
MBC on monogenic amplitude, orientation and phase
could individually achieve high recognition rates. The
fused MBC (MBC-A+MBC-P+MBC-O) could further
improve the recognition accuracy and achieve state-of-
the-art FR performance. MBC has much less time and
space complexity than the widely used Gabor
transformation methods. The demerit of monogenic
representation and face recognition using MBC is complex
in terms of implementation.

2.5 GABOR ORDINAL MEASURES
LBP only considers the qualitative relationship between
two pixels rather than ordinal measures among multiple
image regions. By combining advantages of both Gabor
wavelets and some kinds of ordinal filters, a novel local
feature analysis method named Gabor Ordinal Measures
(GOM) is introduced [9]. It uses di-lobe and tri-lobe
ordinal filters to effectively encode Gabor magnitude and
phase images. The basic idea of GOM is to integrate
distinctiveness of Gabor features and robustness of this
kind of ordinal measures as a promising solution to jointly
handle inter-person similarity and intra-person variations
in face images.
2.5.1 Ordinal Measure
Ordinal measure (OM) is defined as the relative ordering
Information of multiple variables [10]. Ordinal measure
can be used to encode the information in intensity level or
feature level. For example, intensity level ordinal measure
may be the qualitative relationship between the average
intensity values of two image regions. If region A is
brighter than region B, the ordinal measure is “A > B” and
represented by one bit feature code “1”. Otherwise “B < A”
and the ordinal code is “0”. Feature level ordinal measure
is the qualitative information computed on the image
features, e.g. Gabor features.
A straightforward idea of ordinal feature extraction is to
compare the weighted average intensity value of
dissociated image regions. Such a process can be
implemented by ordinal filtering. If the filtering result is
positive, the derived ordinal code is “1”. Otherwise it is “0”.
An ordinal filter is usually constituted by multiple positive
lobes and negative lobes. So Multi-lobe differential filters
(MLDF) is used for ordinal feature extraction
When comparing MLDF with the original LBP operator, it
is more robust to uniform noise. This is shown in Figure 6 .
Ordinal measures from image intensity values have shown
some robustness on face recognition. However, different
from other texture rich biometrics modalities such as iris
pattern, it is difficult to find enough robust and identity
related ordinal relationships in a face image. Firstly, local
intensity variations are not significant in face region
because of the similar reflection ratio of facial skin.
Secondly, the discriminating power of image intensity
ordinal measures is limited because of the similarity
between inter-class facial appearances. A good solution to
this problem is to derive ordinal measures from feature
level, after encoding some of the face information. The use
of Gabor filters first in order to enhance the local details of
face texture.
The benefit is twofold: on the one hand, local variations of
Gabor images are significant so that we can derive more
stable ordinal measures from them; on the other hand,
multi-channel Gabor filters can selectively obtain the
texture information at a given orientation and scale.
TABLE I. RECOGNITION RATE OF DIFFERENT METHODS
ON FERET DATABASE
Method Fb Fc Dup I DupII
LBP 93.0 51.0 61.0 50.0
GV-LBP-TOP-M 98.08 98.45 80.89 81.20
GV-LBP-TOP-P 97.99 98.97 81.86 83.76
E-GV-LBP-M 98.41 98.97 81.99 81.62
E-GV-LBP-P 97.82 97.42 80.47 78.63
MBC-A 98.10 97.40 74.40 72.20
MBC-P 96.60 99.00 81.60 78.20
MBC-O 97.60 99.00 85.20 83.30
MBC-F 99.70 99.50 93.60 91.50
GOM 99.90 100 95.70 93.10
Hence, the face image is decomposed into multiple
spectrum channels and the feature space of ordinal
measures is enlarged to discriminate face images with
inter-class similarity.
The ordinal measures derived from Gabor images are
named Gabor Ordinal Measures (GOM).
The following reasons motivated the proposal of a novel
face recognition method:
1) Ordinal measures are robust image descriptors for face
feature representation.
2) GOM integrates the robustness advantage of ordinal
measures and distinctiveness advantage of Gabor
wavelets, providing a promising solution for face image
analysis.
3) By combining ordinal measures derived from image
intensity level and feature level, a better face recognition
performance can be achieved.
The procedure Gabor Ordinal Measures (GOM) is as
follows.First, Ordinal Measures on Gabor (GOM)
magnitude, phase, real and imaginary responses at
different orientations and scales are extracted. Then, the
ordinal information of multiple orientations is encoded
into GOM maps. Finally, spatial histograms are extracted
from each of the encoded maps. Besides, the use of LDA
makes the representation more compact.
The main contributions of this method are
1), Ordinal measures extracted on feature level are more
discriminative than those extracted on image level.
2) Fusion of image level ordinal measures (GOM-r and
GOM-i) and feature level ordinal measures (GOM-m and
GOM-p) can achieve better recognition performance.
4) Higher accuracy of face recognition can be achieved
with better setting of the GOM parameters.

TABLE II. COMPARISON OF GOM WITH OTHER METHODS
The recognition performances of the aforementioned
methods are summarized in table I. From the table LBP
has lesser Fc value.ie; LBP is more sensitive to noise. GV-
LBP-TOP phase feature provides high recognition rate as
compared to magnitude feature except image with
different facial expression. But in the case of Effective GV-
LBP magnitude information achieves better recognition
performance than phase information for entire probe
sets(fb,fc,dup I and dupII).The MBC-A provides better
recognition rate than MBC-P and MBC-O in different facial
expressions .The fused MBC (MBC-F) achieved better
recognition rate than individual MBC codes. From the
table we can analyse that GOM achieves better recognition
performance than other methods. So it is a promising
solution to jointly handle inter-person similarity and intra-
person variations in face images.
Table II list the advantages and disadvantages of
aforementioned methods.
Methods Advantages Disadvantages
LBP Robust to noise in non uniform region. Degrade the performance due to
changes of lighting direction and
shadowing.
Sensitive to noise in uniform
regions.
Gabor wavelet Extract features from multiple orientation and scale. High dimensionality of feature
vector.
high computational cost and
storage space
LTP More discriminative and less sensitive to noise in
uniform regions.
Not strictly invariant to grey level
transformation.
LBP and Gabor
wavelet
a)GV-LBP-TOP Discriminative and robust information is extracted from
three domains.
High computational complexity.
b) E-GV-LBP Reduced computational cost.
MBC Higher accuracy and low cost.
Less time and space complexity than other Gabor based
methods.
Complex in terms of
implementation.
GOM Promising solution for face image analysis.
Better performance and higher accuracy.
Matching process is highly efficient.
High computational complexity.

3. CONCLUSIONS
In this paper various local feature based approaches are
discussed. Two representative methods of local feature
analysis in face biometric are Gabor wavelet and LBP.LBP
is basically a fine descriptor which captures small texture
details, in contrast to Gabor feature, which encodes the
facial shape and appearance over a range of scales. The
process of extracting LBP features is much faster than that
of Gabor based feature. The LBP is sensitive to noise in
uniform regions. So a generalization of LBP, Local Ternary
Pattern (three valued code) is introduced.
To improve the FR performance both Gabor wavelet and
LBP are combined.GV-LBP-TOP is one such method in
which information are obtained by assembling scale,
orientation and spatial domain to form third order
volume. This domain provides complementary
information, which is useful for FR. Shortcoming of GV-
LBP-TOP is the high computational complexity. To address
this problem Effective GV-LBP (E-GV-LBP) is introduced.
This method encodes the information in spatial, frequency
and orientation domain simultaneously and reduces
computational cost. One of the shortcomings of Gabor
wavelets is that it is computationally expensive in terms of
cost and storage. So MBC is introduced, which achieves
less time and space complexity than the Gabor based
methods. Then, finally GOM is provided which
outperforms all the method mentioned here .The idea of
GOM is to integrate distinctiveness of Gabor features and
robustness of this kind of ordinal measures as a promising
solution to jointly handle inter-person similarity and intra-
person variations in face images.
REFERENCES
[1] M. A. Turk and A. P. Pentland,“Face recognition using
eigenfaces,” in Proc. IEEE Comput. Soc. Conf. Comput. Vis.
Pattern Recognit., Jun. 1991, pp. 586–591.
[2]P.Belhumeur,J.Hespanha, and D.Kriegman,“Eigenfaces
vs. fisher faces: Recognition using class specific linear
projection,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 19,
no. 7, pp. 711–720, Jul. 1997.
[3]P. Comon, “Independent component analysis—a new
concept?” Signal Process., vol. 36, pp. 287–314, 1994.
[4] Á. Serrano, I. M. de Diego, C. Conde, and E. Cabello,
“Recent advances in face biometrics with Gabor wavelets:
A review,” Pattern Recognit. Lett., vol. 31, no. 5, pp. 372–
381, 2010.
[5]T.Ojala,M.Pietikäinen,andT. Mäenpää,“Multiresolution
gray-scale and rotation invariant texture classification
with local binary patterns,”IEEE Trans. Pattern Anal. Mach.
Intell., vol. 24, no. 7, pp. 971–987, Aug. 2002.
[6]X. Tan and B. Triggs, “Enhanced local texture feature
sets for face recognition under difficult lighting
conditions,” IEEE Trans. Image Process., vol. 19, no. 6, pp.
1635–1650, Jun. 2010.
[7] Z. Lei, S. Liao, M. Pietikainen, and S. Li, “Face
recognition by exploring information jointly in space, scale
and orientation,” IEEE Trans. ImageProcess., vol. 20, no. 1,
pp. 247–256, Jan. 2011.
[8]M. Yang, L. Zhang, S. C.-K. Shiu, and D. Zhang,
“Monogenic binary coding: An efficient local feature
extraction approach to face recognition,” IEEE Trans. Inf.
Forensics Security, vol. 7, no. 6, pp. 1738–1751, Dec. 2012.
[9] Zhenhua Chai, Zhenan Sun, Heydi Méndez-Vázquez,
Ran He, and Tieniu Tan “Gabor Ordinal Measures for Face
Recognition” IEEE Trans. Information forensics and
security, vol. 9, no. 1, jan 2014.
[10]S. Liao, Z. Lei, X. Zhu, Z. Sun, S. Z. Li, “Face recognition
using ordinal features,” in Proc. ICB, 2006, pp. 40–46.

A Survey on Local Feature Based Face Recognition Methods

More Related Content

What's hot (20)

Similar to A Survey on Local Feature Based Face Recognition Methods (20)

More from IRJET Journal (20)

Recently uploaded (20)

A Survey on Local Feature Based Face Recognition Methods