See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/4232845

Local Binary Patterns as an Image

Preprocessing for Face Authentication

Conference Paper · May 2006

DOI: 10.1109/FGR.2006.72 · Source: IEEE Xplore

CITATIONS READS

125 66

3 authors, including:

Sébastien Marcel
Idiap Research Institute
147 PUBLICATIONS 2,090 CITATIONS

SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Sébastien Marcel
letting you access and read them immediately. Retrieved on: 22 April 2016
R E P O R T
R E S E A R C H

Local Binary Patterns as an

Image Preprocessing for
Face Authentication
Guillaume Heusch a Yann Rodriguez a

Sébastien Marcel a
IDIAP–RR 05-76
I D I AP

November 2005

a IDIAP Research Institute

IDIAP Research Institute www.idiap.ch

Rue du Simplon 4 P.O. Box 592 1920 Martigny − Switzerland
Tel: +41 27 721 77 11 Fax: +41 27 721 77 12 Email: [email protected]
 IDIAP Research Report 05-76

Local Binary Patterns as an Image

Preprocessing for Face Authentication

Guillaume Heusch Yann Rodriguez Sébastien Marcel

November 2005

Abstract. One of the major problem in face authentication systems is to deal with variations in
illumination. In a realistic scenario, it is very likely that the lighting conditions of the probe image
does not correspond to those of the gallery image, hence there is a need to handle such variations.
In this work, we present a new preprocessing algorithm based on Local Binary Patterns (LBP): a
texture representation is derived from the input face image before being forwarded to the classifier.
The efficiency of the proposed approach is empirically demonstrated using both an appearance-
based (LDA) and a feature-based (HMM) face authentication systems on two databases: BANCA
and XM2VTS (with its darkened set). Conducted experiments show a significant improvement in
terms of verification error rates and compare to results obtained with state-of-the-art preprocessing
techniques.
2 IDIAP–RR 05-76

1 Introduction
Since over 30 years, face recognition has been an active research area and different systems are now
capable to properly recognize people under controlled environment (i.e. when the probe and the
gallery images are acquired under the same global viewing conditions) [24]. Nevertheless, there are
still open challenges such as recognition with varying pose, expression or illumination. In this paper,
we will focus on the latter. Existing algorithms to cope with changes in lighting conditions can be
divided into two main categories: model-based and preprocessing.
Model-based approaches derive a model of an individual face, which will account for variation
in lighting conditions. Such algorithms relies on the assumption that the collection of face images
under any illumination condition (the so-called face space) is a low-dimensional subspace of the image
space. Several algorithms were developed based on this idea: examples include the spherical harmonics
representation [2] and the illumination cone [8]. However, such algorithms require a training set with
several different lighting conditions for the same subject, which can be considered as a weakness for
realistic applications.
On the other hand, the preprocessing approach tries to minimize the effect of illumination varia-
tions prior to classification. Compared to the first approach, preprocessing has two main advantages:
it is completely stand-alone and thus can be used with any classifier. Moreover, no training images are
required to model illumination variations, since such algorithms turn the original image into another
representation, which is supposed to be less sensitive to such variations. A recent study of preprocess-
ing algorithms applied to face authentication [19] showed the superiority of the model proposed by
Gross & Brajovic [9]. However, this approach involves the resolution of a partial differential equation,
which is either computationaly expensive (brute force resolution) or approximate (iterative methods).
Moreover, the algorithm depends on a normalization parameter and hence face authentication cannot
be performed in a fully automated way, since this parameter has to be manually set a priori.
In this work, we propose to use Local Binary Patterns (LBP) [17] as a preprocessing step prior
to the face authentication task. The LBP is a nonparametric and computationnaly simple descriptor
of local texture patterns. Moreover, it is invariant to monotonic grayscale transformation, hence the
LBP representation may be less sensitive to changes in illumination.
The remainder of this paper is organized as follows: in the next section, we briefly describe the
normalization proposed in [9] and present our approach based on LBP. Section 3 introduces the task
of face authentication and presents the systems we used in our experiments. Experimental setup as
well as authentication results are presented in section 4 and a conclusion is drawn in section 5.

2 Illumination Normalization
The aim of this section is to describe the preprocessing steps we investigate in our experiments. As
a baseline, the well-known histogram equalization was used prior to classification. The remainder of
the section presents the state-of-the-art algorithm developed by Gross & Brajovic [9], as well as the
approach we propose, based on Local Binary Patterns.

2.1 Gross & Brajovic Model

In their paper [9], Gross and Brajovic addressed the problem of lighting normalization by using calculus
of variations. Recovering the reflectance from the illuminance-reflectance model, where an image is
regarded as the product between the illuminance and the reflectance: I(x, y) = R(x, y)I(x, y), is an
ill-posed problem. Based on the common assumption that the luminance L(x, y) can be estimated
as a blurred version of the original image, they derive the following energy-based model to recover
L(x, y):
ZZ ZZ
2
E(L) = ρ(x, y)(L − I)) dxdy + λ (L2x + L2y )dxdy (1)
IDIAP–RR 05-76 3

where the first term forces the luminance function to be close to the image and the second term adds
a smoothness constraint on L(x, y). ρ(x, y) are the anisotropic diffusion coefficients and the parameter
λ controls the relative importance of the smoothness constraint. Note also that if ρ(x, y) = 1 ∀ x, y,
then isotropic diffusion is performed. The minimum of this functional is found using the corresponding
discretized Euler-Lagrange equation:

·
1 1
Li,j +λ ρi,j− (Li,j − Li,j−1 ) + − Li,j+1 )
ρi,j+ (Li,j
¸
1 1
+ ρi−,j (Li,j − Li−1,j ) + ρi+,j (Li,j − Li+1,j ) = Ii,j

(2)
1
where ρa,b denotes the anisotropic coefficients and are defined as the inverse of Weber’s contrast
(equation 3). The resolution of equation (2) is carried out using iterative methods [11].
|Ia − Ib |
ρa,b = (3)
min(Ia , Ib )

2.2 Our Approach: Local Binary Patterns

The local binary pattern (LBP) is a non-parametric operator which describes the local spatial structure
of an image. Ojala et al. [17] first introduced this operator and showed its high discriminative power
for texture classification. At a given pixel position (xc , yc ), LBP is defined as an ordered set of binary
comparisons of pixel intensities between the center pixel and its eight surrounding pixels (Figure 1).
The decimal form of the resulting 8-bit word (LBP code) can be expressed as follows:
7
X
LBP (xc , yc ) = s(in − ic )2n (4)
n=0

where ic corresponds to the grey value of the center pixel (xc , yc ), in to the grey values of the 8
surrounding pixels, and function s(x) is defined as:
½
1 if x ≥ 0
s(x) = (5)
0 if x < 0 .

By definition, the LBP operator is unaffected by any monotonic gray-scale transformation which

binary
83 75 126 0 0 1
intensity
binary: 00111001
99 95 141 1 1
decimal: 57
comparison
91 91 100 0 0 1
with the center

Figure 1: The LBP operator.

preserves the pixel intensity order in a local neighbourhood. Note that each bit of the LBP code has
the same significance level and that two successive bit values may have a totally different meaning.
Actually, The LBP code may be interpreted as a kernel structure index.
Later, Ojala et al. [18] extended their original LBP operator to a circular neighbourhood of different
radius size. Their LBPP,R notation refers to P equally spaced pixels on a circle of radius R. In this
paper, we use the LBP8,2 operator which is illustrated in Figure 2.
Due to its texture discriminative property and its very low computational cost, LBP is becoming
very popular in pattern recognition. Recently, LBP has been applied for instance to face detection [12],
4 IDIAP–RR 05-76

Figure 2: The extended LBP operator with (8,2) neighbourhood. Pixel values are interpolated for
points which are not in the center of a pixel.

face recognition [23, 1], image retrieval [21], motion detection [10] or visual inspection [22]. In these
previous works an image was usually divided in small regions. For each region, a cumulative histogram
of LBP codes, computed at each pixel location within the region, was used as a feature vector.
However, in the present work, we do not use LBP to perform feature extraction, but to preprocess
the input image: the face is represented with its texture patterns given by the LBP operator at each
pixel location (Figure 3).

LBP operator

Figure 3: Original image (left) processed by the LBP operator (right).

3 Face Authentication
The task of face authentication is to make a decision, based on the face image of a client C and
on a claimed identity: either the claimed identity corresponds to the identity derived from the face
image (true claimant), or the client is an impostor. Hence, a face authentication system has to give
an opinion on whether the client is the true claimant or not, given its face image. In this section,
we present the two different face authentication algorithms used in our experiments. The first one is
based on linear discriminant analysis (LDA) and is said to be holistic: the whole face is treated as a
single vector. On the contrary, the second one is based on local features and Hidden Markov Models
(HMM).

3.1 PCA-LDA
Linear discriminant analysis was first applied in face recognition in [3] and has been successfully
applied to the problem of face authentication [13] [14]. The LDA algorithm is based on Fisher’s linear
discriminant formula to find the optimal (discriminative) subspace where data will be projected. Let
us consider a training set composed of N examples xi where each example is labeled and belongs to
one of C classes. The training data are used to derive the between-class scatter matrix S b and the
within-class scatter matrix Sw :

C
X
Sb = Ni (µi − µ)(µi − µ)T (6)
i=1
C
X X
Sw = (xk − µi )(xk − µTi ) (7)
i=1 xk ∈Xi
IDIAP–RR 05-76 5

where Ni are the number of examples in class i, µi is the mean of the examples in class i and µ is
the mean of all training data. Finally, the optimal projection Wlda is the orthogonal matrix satisfying
equation (8):

|W T Sb W |
Wlda = argmax (8)
W |W T Sw W |
In practice, Principal component analysis (PCA) is first applied on the data so as to achieve
decorrelation and dimensionality reduction. The projected face representations into the coordinate
system of eigenvectors are then used as training examples to derive the optimal projection W lda .

Authentication Considering the holistic face representation as a vector denoted by x, it is linearly

projected to obtain the feature vector y:
T T
y = Wlda Wpca x

T
where Wpca is the PCA matrix, containing the eigenvectors of the training data covariance matrix.
Finally, classification is performed using a metric in the feature space: considering two feature vectors,
a template yt and a sample ys , their correlation is computed according to the following formula:
< y t , ys >
corr (ys , yt ) = 1 −
kyt k · kys k

The decision on the identity claim is based on a threshold τ : the claim is accepted if corr (y s , yt ) ≤ τ
and rejected otherwise.

3.2 1D-HMM
It has been shown that face authentication systems based on statistical models are suitable for the
face authentication task [6]. On the contrary to the LDA-based system where a face image is treated
as a whole, such models generally uses local features to analyze a given face image. Several statistical
models have been proposed to model the face representation, examples include Gaussian Mixtures
Models (GMM) [6], Hidden Markov Models (HMM) [5] and Bayesian networks [16].
With the one dimensional HMM (1D-HMM) structure proposed in [4] the face representation is
analyzed on a block-by-block basis: from each block, discrete cosine transform (DCT) coefficients
are obtained; the first M coefficients compose the feature vector corresponding to the block under
consideration. The main advantage of this topology is to relax the rigid horizontal constraints present
in the traditional 1D-HMM, where an entire strip is used as observation.
Such a model is characterised by its number of states Ns , its state transition matrix A and its
emission probability distributions B, which are considered to be a mixture of gaussians. Let us define
X as the observation sequence, representing the consecutive horizontal strips of the face: X = {x s }N
s=1 .
s

NB
Each strip is itself a sequence of NB observation vectors xs = {xb }b=1 representing the consecutive
blocks composing the strip. Under the assumption that the blocks composing a strip are independent
and identically distributed, the likelihood of the strip s for the state Sj is given by:

NB
Y
P (~xs |qs = Sj ) = P (~xbs |λj ) (9)
b=1

Then, the likelihood of an observation sequence X is:

NS Y
XY NB NS
Y
P (X|λ) = P (~xbs |λj ) aqs−1 ,qs (10)
∀Q s=1 b=1 s=2
6 IDIAP–RR 05-76

where:
NG
X
P (~xb |λj ) = µsk , Σsk )
msk N (~xbs |~ (11)
k=1

The training phase is performed in two stages: first, a so-called world-model is learned using image
data coming from various identities. Then, each client model is adapted from the world-model through
maximum a posteriori (MAP) adaptation [5].

Authentication Let us denote λC̄ as the parameter set defining the world-model whereas λC rep-
resents the client-specific parameters. Given a client claim and its face representation X, an opinion
on the claim is given by the following likelihood ratio:

Λ(X) = log P (X|λC ) − log P (X|λC ) (12)

where P (X|λC ) is the likelihood of the claim coming from the true client and P (X|λC ) is an approx-
imation of the likelihood of the claim coming from an impostor. Based on a threshold τ , the claim is
accepted if Λ(X) ≥ τ and rejected otherwise.

4 Experiments & Results

Face windows are manually extracted from the original images based on the eyes position. They are
resized to 64x80 pixels, converted to grayscale and finally treated by the investigated preprocessing
algorithms before training and testing. Tables 1 and 2 presents results for each face authentication
systems with three preprocessing techniques: histogram equalization (denoted as HEQ), the normal-
ization model proposed in [9] (denoted as GROSS) and our approach (denoted as LBP). Performance
are reported in terms of Half Total Error Rate (HTER), which combines the False Acceptance Rate
and the False Rejection Rate:
(F AR + F RR)
HT ER = [%] (13)
2

4.1 XM2VTS
The XM2VTS database [15] is a multi-modal database containing 295 idendities, among which 200
are used as true clients (the remainder are considered as impostors). Recordings were acquired during
four sessions over a period of five months under controlled conditions (blue background, uniform
illumination) for the standard set. The darkened set contains four images of each subject acquired
with side lighting. Figure 4 shows images coming from both sets of the database.
Together with the standard set, there exists two detailed protocols used to perform authentication
experiments [15]. Concerning darkened set experiments, protocols are the same but for the testing
phase: it is done on the darkened images. Table 1 shows the results obtained on the XM2VTS
database, where LP1 and LP2 refers to the Lausanne Protocol 1 and 2 respectively.

4.2 BANCA
The BANCA database [7] was especially meant for multi-modal biometric authentication and contains
52 clients (English corpus), equally divided into two groups g1 and g2 used for development and eval-
uation respectively. Each corpus is extended with an additional set of 30 other subjects and is referred
as the world model. Image acquisition was performed with two different cameras: a cheap analogue
webcam, and a high-quality digital camera, under several realistic scenarios: controlled (high-quality
camera, uniform background, controlled lighting), degraded (webcam, non-uniform background) and
adverse (high-quality camera, arbitrary conditions). Figure 5 shows examples of the different acqui-
sition scenarios.
IDIAP–RR 05-76 7

(a) standard set (b) darkened set

Figure 4: examples of images in the XM2VTS database.

standard darkened
FA system LP1 LP2 LP1 LP2
LDA HEQ 2.97 0.84 10.86 17.02
LDA GROSS 5.76 4.88 12.62 13.38
LDA LBP 4.56 1.43 9.110 10.44
HMM HEQ 2.04 1.40 37.32 37.54
HMM GROSS 5.53 4.18 12.01 11.96
HMM LBP 1.37 0.97 9.61 9.88

Table 1: Performances on the standard and the darkened sets of the XM2VTS database.

(a) controlled (b) degraded (c) adverse

Figure 5: Example of the different scenarios in the BANCA database

In the BANCA protocol, seven distinct configurations for the training and testing policy have
been defined. In our experiments, the configurations referred as Match Controlled (Mc), Unmatched
Adverse (Ua), Unmatched Degraded (Ud) and Pooled Test (P) are used. All of the of listed configu-
rations use the same training conditions: each client is trained using images from the first recording
session, which corresponds to the controlled scenario. Testing is then performed on images taken from
the controlled scenario (Mc), adverse scenario (Ua), degraded scenario (Ud), while (P) does the test
for each of the previously described configurations. Table 2 shows the results obtained on the different
protocols of the BANCA database.
8 IDIAP–RR 05-76

FA system Mc Ua Ud P
LDA HEQ 3.75 20.13 14.46 15.52
LDA GROSS 3.97 17.40 15.01 14.24
LDA LBP 5.83 19.52 15.61 16.30
HMM HEQ 2.40 19.87 18.75 18.32
HMM GROSS 1.92 11.70 7.21 11.75
HMM LBP 2.40 15.06 9.93 11.70

Table 2: Performance on the different protocols of the BANCA database.

4.3 Discussion
Firstly, results on the XM2VTS database show that the LBP representation is suitable when there
is a strong mismatch, in terms of illumination conditions, between the gallery and the probe image.
This is evidenced by experiments on the darkened set, where the error rates of both classifiers are
decreased when using the LBP representation. Moreover, this texture representation outperforms
the illumination normalization approach (GROSS). Interestingly, standard experiments also show an
improvement for the HMM-based classifier: this suggest that our preprocessing technique is well
suited for feature-based approaches. Although the best results obtained with LDA are with the use
of histogram equalization, error rates of the LBP are still lower than the GROSS normalization.
On the BANCA database, the LDA classifier seems to have a good discriminative capability, since
none of the method clearly outperforms the others (although GROSS normalization is the best). A
possible explanation could reside in the fact that we use the Spanish corpus (with all scenarios) to
train the LDA, hence it may capture by itself the changes in acquisition conditions. Concerning
the HMM-based classifier, GROSS normalization results are better for three of the four investigated
protocol, and reduces error rates by a significant amount compared to histogram equalization. Results
obtained with the LBP representation are comparable, although performances are a bit worse.
To summarize, conducted experiments shows that the proposed preprocessing approach is suitable
for face authentication: results are comparable with, or even better than those obtained using the
state-of-the-art preprocessing algorithm proposed in [9]. Moreover, the LBP representation has several
other advantages: there is no need for hyperparameter selection, hence avoiding extensive experiments.
Indeed, the normalization results were obtained with different values of λ, depending on the database
and on the classifier (a recent study [20] shows that better results are obtained if the normalization
parameter is tuned for each probe image) whereas the LBP preprocessing was always the same.
Furthermore, the LBP operator is simpler and faster than the illumination normalization process.

5 Conclusion
In this paper, we proposed an original preprocessing technique based on LBP for illumination robust
face authentication. The LBP operator provides a texture representation of the face which improves
the performances of two different classifiers (holistic and feature-based) as compared to histogram
equalization. Moreover, obtained results are comparable when using the state-of-the-art preprocessing
algorithm [9] on the same databases [20], and removes the need for parameter selection.

References
[1] T. Ahonen, A. Hadid, and M. Pietikäinen. Face recognition with local binary patterns. In
European Conference on Computer Vision (ECCV), 2004.
IDIAP–RR 05-76 9

[2] R. Basri and D.W. Jacobs. Lambertian reflectance and linear subspaces. IEEE Trans. on Pattern
Analysis and Machine Intelligence, 25(2):218–233, 2003.

[3] Peter N. Belhumeur, Joao Hespanha, and David J. Kriegman. Eigenfaces vs. fisherfaces: Recog-
nition using class specific linear projection. IEEE Trans. on Pattern Analysis and Machine
Intelligence, 19(7):711–720, 1997.

[4] F. Cardinaux. Local Features and 1D-HMMs for Fast and Robust Face Authentication. IDIAP-
RR 17, IDIAP, 2005.

[5] F. Cardinaux, C. Sanderson, and S. Bengio. Face Verification Using Adapted Generative Models.
In IEEE Automatic Face and Gesture Recognition (AFGR), 2004.

[6] F. Cardinaux, C. Sanderson, and S. Marcel. Comparison of MLP and GMM classifiers for face ver-
ification on XM2VTS. In Audio- and Video-based Biometric Person Authentication, AVBPA’03,
2003.
[7] E. Bailly-Baillière et al. The banca database and evaluation protocol. In Audio- and Video-based
Biometric Person Authentication, AVBPA’03, 2003.

[8] A. Georghiades, P. Belhumeur, and D. Kriegman. From few to many: Illumination cone models
for face recognition under variable lighting and pose. IEEE Trans. on Pattern Analysis and
Machine Intelligence, 23(6):643–660, 2001.

[9] R. Gross and V. Brajovic. An image preprocessing algorithm for illumination invariant face
recognition. In Audio- and Video-Based Biometric Person Authentication, AVBPA’03, 2003.

[10] M. Heikkilä, M. Pietikäinen, and J. Heikkilä. A texture-based method for detecting moving
objects. In British Machine Vision Conference (BMVC), 2004.

[11] G. Heusch, F. Cardinaux, and S. Marcel. Efficient diffusion-based illumination normalization for
face verification. IDIAP RR 46, IDIAP, 2005.

[12] H. Jin, Q. Liu, H. Lu, and X. Tong. Face detection using improved LBP under bayesian frame-
work. In Third International Conference on Image and Graphics (ICIG), Hong Kong, China,
pages 306–309, 2004.
[13] J. Kittler, Y. Li, and J. Matas. On matching scores for lda-based face verification. In British
Machine Vision Conference (BMVC), 2000.
[14] S. Marcel. A symmetric transformation for lda-based face verification. In IEEE Automatic Face
and Gesture Recognition (AFGR), 2004.
[15] K. Messer, J. Matas, J. Kittler, J. Lüttin, and G. Maitre. XM2VTSDB: The extended M2VTS
database. In Audio- and Video-based Biometric Person Authentication, AVBPA’99, 1999.
[16] A. Nefian. Embedded bayesian networks for face recognition. In IEEE International Conference
on Multimedia, ICME, 2002.
[17] T. Ojala, M. Pietikäinen, and D. Harwood. A comparative study of texture measures with
classification based on featured distributions. Pattern Recognition, 29(1):51–59, 1996.
[18] T. Ojala, M. Pietikäinen, and T. Mäenpää. Multiresolution gray-scale and rotation invariant
texture classification with local binary patterns. IEEE Trans. on Pattern Analysis and Machine
Intelligence, 24(7):971–987, 2002.

[19] J. Short, J. Kittler, and K. Messer. A comparison of photometric normalisation algorithms for
face verification. In IEEE Automatic Face and Gesture Recognition (AFGR), 2004.
10 IDIAP–RR 05-76

[20] J. Short, J. Kittler, and K. Messer. Photometric normalisation for face verification. In Audio-
and Video-based Biometric Person Authentication, AVBPA’05, 2005.

[21] V. Takala, T. Ahonen, and M. Pietikäinen. Block-based methods for image retrieval using local
binary patterns. In Proc. 14th Scandinavian Conference on Image Analysis (SCIA), Joensuu,
Finland, pages 882–891, 2005.

[22] M. Turtinen, M. Pietikäinen, and O. Silven. Visual characterization of paper using isomap and
local binary patterns. In Conference on Machine Vision Applications (MVA), 2005.

[23] G. Zhang, X. Huang, S.Z. Li, Y. Wang, and X. Wu. Boosting local binary pattern (LBP)-based
face recognition. In Chinese Conference on Biometric Recognition, SINOBIOMETRICS 2004,
2004.

[24] W. Zhao, R. Chellappa, P.J. Phillips, and A. Rosenfeld. Face recognition: A literature survey.
ACM Computing Surveys, 35(4):399–458, 2003.