iFR: Interactively Pose Corrected Face Recognition
Simon Nash, Mark Rhodes and Joanna Isabelle Olszewska
School of Computing and Technology, University of Gloucestershire, The Park, Cheltenham, GL50 2RH, U.K.
Keywords:
Face Detection, Face Recognition, AdaBoost, Template, Facial Features, Eigenfaces, Pose Correction,
Human-Computer Interactive Systems.
Abstract:
Although face recognition applications are growing, robust face recognition is still a challenging task due
e.g. to variations in face poses, facial expressions, or lighting conditions. In this paper, we propose a new
method which allows both automatic face detection and recognition and incorporates an interactive selection
of facial features in conjunction with our new pose-correction algorithm. Our resulting system we called iFR
successfully recognizes faces across pose, while being computationally efficient and outperforming standard
approaches, as demonstrated in tests carried out on publicly available standard datasets.
1 INTRODUCTION
In recent years, computational face recognition has at-
tracted an increasing attention because of its potential
use in biometric systems (Jain et al., 2012), video-
surveillance (Rambach et al., 2015), automated be-
havior analysis (Han et al., 2013), social networks
(Choi et al., 2011), or human-machine interactions
(Tasli et al., 2015).
For this purpose, different techniques for auto-
matic face recognition have been developed. These
methods could rely on facial features (Brunelli and
Poggio, 1993), active appearance models (Cootes
et al., 2001), active shape models (Lindner et al.,
2015), or facial representations like Local Binary Pat-
terns (LBP) (Ahonen et al., 2006), (Montagne et al.,
2013). Other approaches use face decomposition in
eigenfaces (Turk and Pentland, 1991) or empirical
mode decomposition (EMD) (Gallego-Jutgla et al.,
2013). Face recognition techniques could also involve
neural networks (Nagi et al., 2008).
Despite these methods have been proven to be ef-
ficient, many challenges remain. In particular, pose
variation has been identified as of great impact on
the face recognition scores (Du and Ward, 2006). In-
deed, face appearance and/or shape could change due
to head’s movements which can be described by the
egocentric rotation angles, namely, pitch, roll, and
yaw (Murphy-Chutorian and Trivedi, 2009), leading
among others to intra-subject variations and mak-
ing the face recognition across pose a task difficult
(Zhang and Gao, 2009).
In the literature, few works have proposed to study
face recognition under pose variations. In (Bordoloi
and Sarma, 2009), they applied a pose correction by
re-enforced multi-level decision support, using a gen-
eralized feed forward artificial neural network (GF-
FANN) designed to detect the rotation angle of an
input image and to correct it by applying the three
pass shear rotation. (Nakamura and Takano, 2011)
developed a rotation and size spreading associative
neural network (RS-SAN), while (Huang et al., 2012)
adopted a deep learning approach, using a deep belief
network (DBN). These methods have performance in
high 80% at extensive computational cost of dozens
of seconds, whereas (Bordoloi and Sarma, 2009)
achieve recognition success rates in 90% within sev-
eral hundreds of seconds.
In this work, we propose to tackle with pose vari-
ations by introducing a new algorithm for automatic
pose correction in order to efficiently process face
recognition. Our algorithm makes decision based on
feature points such as the nose tip and the nose top
rather than on block facial features as in (Bordoloi
and Sarma, 2009).
Our face recognition method deals with recogni-
tion of faces in arbitrary, in-depth rotations.
Our iFR system provides both flexibility in the
choice of facial points by integrating a user-driven, in-
teractive phase, and automation in the phase of pose
correction through the new algorithm. On the other
hand, our face recognition approach combines advan-
tages of both facial features and template approaches
as demonstrated in Section 3.
106
Nash, S., Rhodes, M. and Olszewska, J.
iFR: Interactively Pose Corrected Face Recognition.
DOI: 10.5220/0005857801060112
In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 4: BIOSIGNALS, pages 106-112
ISBN: 978-989-758-170-0
Copyright
c
2016 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
Figure 1: Example of our iFR system in action, featuring
automatic face detection, interactive pose correction, and
automatic face recognition.
The contribution of this paper is twofold. It con-
sists, on one hand, in a new face recognition system
which enables a user to select interactively facial fea-
ture points, such as nose tip and nose top, in order to
feed the automatic process correcting the face pose.
On the other hand, the pose correction itself is done
by a new algorithm which we present in this paper and
which contributes to improve face recognition perfor-
mance.
The paper is structured as follows. In Section
2, we describe our interactively pose corrected face
recognition method (see Fig. 1). The resulting sys-
tem, that automatically detects and recognizes a face
which pose has been corrected by our algorithm, has
been successfully tested on standard dataset, as re-
ported and discussed in Section 3. Conclusions are
drawn up in Section 4.
2 PROPOSED APPROACH
Our pose-corrected face recognition (iFR) system
consists of three phases. At first, the automatic face
detection, described in Section 2.1, performs the de-
tection and location of a face in an image. Secondly,
the located face is cropped automatically and serves
as input to the second step. This latter one is the inter-
active phase of the system as explained in Section 2.2.
Indeed, it is the user selecting two facial feature points
based on which the pose correction is computed auto-
matically. Thirdly, the recognition phase processes
that pose-corrected face image, and provides the sys-
tem output which could be either the recognized face
or the ‘unrecognized pictogram’, as detailed in Sec-
tion 2.3.
2.1 Automatic Face Detection
In our iFR system, the first phase aims to automati-
cally detect and locate face(s) in a given image I(x, y)
with M and N, its width and height, respectively. For
this purpose, we implemented the Viola & Jones ap-
proach (Viola and Jones, 2004), which is based firstly
on the detection of Haar-like features and secondly on
the processing of the AdaBoost classifier.
Local Haar-like features f encode the existence of
oriented contrasts between regions in the processed
image. They are computed by subtracting the sum of
all the pixels of a subregion of the local feature from
the sum of the remaining region of the local feature,
using the integral image representation II(i, j) which
is defined as follows:
II(i, j) = II(i 1, j ) + II(i, j 1) II(i 1, j 1) + I(i, j),
(1)
where I(i, j) is the pixel value of the original image at
the position (i, j) (Wood and Olszewska, 2012).
The AdaBoost classifier requires at first a training
phase during which it builds strong classifiers based
on cascades of weak classifiers. In particular, to form
a T -stage cascade, T weak classifiers are selected us-
ing the AdaBoost algorithm (Viola and Jones, 2004).
In fact at a t stage of this cascade, a sub-window u of
an image from the training set is computed by eq. (1)
and it is passed to the corresponding t
th
weak classi-
fier. If the region is classified as a non-face, the sub-
window is rejected. If not, it is passed to the t + 1
stage, and so forth. Consequently, more stages the
cascade owns, more selective it is, i.e. less false posi-
tive detections occur. However, that could lead to the
increase in the number of false negative detection.
In order to select at each t level (with 1 < t < T )
the best weak classifier, an optimum threshold θ
t
is
computed by minimizing the classification error due
to the selection of a particular Haar-like feature value
f
t
(u). The resulting weak classifier k
t
is thus obtained
as follows:
k
t
(u, f
t
, p
t
, θ
t
) =
(
1 if p
t
f
t
(u) < p
t
θ
t
,
0 otherwise,
(2)
where p
t
is the polarity indicating the direction of the
inequality.
Then, a strong classifier K
T
(u) is constructed by
taking a weighted combination of the selected weak
classifiers k
t
according to
K
T
(u) =
(
1 if
T
t=1
α
t
k
t
(u)
1
2
T
t=1
α
t
,
0 otherwise,
(3)
where
1
2
T
t=1
α
t
is the AdaBoost threshold and α
t
is
a voting coefficient computed based on the classifica-
tion error in each stage t of the T stages of the cascade
(Viola and Jones, 2004).
iFR: Interactively Pose Corrected Face Recognition
107
Once the training phase is completed, the Ad-
aBoost classifier could be applied to a test image. The
Viola & Jones’ face detector (Viola and Jones, 2004)
outputs then a bounding box surrounding the detected
face (Wood and Olszewska, 2012), rather than its con-
tour (Olszewska, 2013). Next, this bounding box’ size
and location are used to automatically crop out the
face from the image, and our iFR system saves it as a
new file.
2.2 Interactive Pose Correction
This phase of pose correction consists of two steps,
one interactive and one automatic.
Hence, once the cropped face is uploaded into the
iFR system (see bottom left of Fig. 1), the user at first
determines the nose tip point (see Fig. 2(a)) and then
the nose top point (see Fig. 2(b)) interactively. For
that, the user clicks each time on the displayed face
image to determine the nose tip point and the nose
top point, respectively.
(a) (b)
Figure 2: Interactive user’s selection of the (a) nose tip; (b)
nose top.
Secondly, these data are passed onto our system
algorithm (Algorithm 1) in order to correct the face
pose automatically.
Algorithm 1 performs first a flip of the face image,
if need be (Fig. 3(a)). Indeed, in first instance, it com-
pares the coordinates of the nose tip identified by the
user with the value of half the height of the cropped
image in order to determine whether the nose point is
higher or lower than this point. If the nose tip is lower
than the halfway line, then the image is already the
right way up. However, if the nose tip is higher than
the halfway line, the face is upside down, and then, is
flipped 180
to turn the face the right way up.
Next, Algorithm 1 computes the angle to rotate the
face in order to correct its pose (Fig. 3(b)). Hence, it
automatically identifies the center point of the image,
while the center point between both eyes, or node top,
has been interactively selected by the user. Then, it
calculates, on one hand, the segment line1 between
the center point of the image and the nose top, and,
(a)
(b)
Figure 3: Illustration of the automatic pose correction
method steps: (a) face flip; (b) face rotation. Best viewed in
color.
on the other hand, the segment line2 between the
nose top and its projection on the vertical halfway
line. Next, the lengths of the segments are computed
as well as the in-between angle using trigonometric
rules. Finally, the image is rotated according to that
angle for pose correction and saved in a file to be used
for automatic face recognition.
2.3 Automatic Face Recognition
The automated face recognition phase is a one-to-
many process (Du and Ward, 2006). It compares a
specific face image (probe) against all of the images
in a face database (gallery). In particular, iFR sys-
tem uses the eigenface approach (Turk and Pentland,
1991), which is based on a holistic region-based rep-
resentation and pixel intensity features, and which in-
troduces the principal component analysis (PCA) of
the distribution of faces, or eigenvectors of the covari-
ance matrix of the set of face images. These eigen-
vectors can be considered as a set of features which
together characterize the variation between face im-
ages. Each image location contributes more or less to
each eigenvector. So, an eigenvector can be seen as a
sort of ghostly face called eigenface. Probe face clas-
sification, i.e. recognition of a test face, is then deter-
mined by applying the nearest neighbour approach to
the probe face projection in face space.
Technically, eigenface-based face recognition is
split in a training step and a testing step.
BIOSIGNALS 2016 - 9th International Conference on Bio-inspired Systems and Signal Processing
108
(a) (b)
Figure 4: Example of the training dataset used for face recognition: (a) initial training dataset; (b) normalized training dataset.
Algorithm 1: Automatic Pose Correction.
Given F = (F
x
, F
y
), the cropped face image,
with its center point c = (HF
x
, HF
y
);
Given the selected nose tip point p = (NP
x
, NP
y
);
Given the selected nose top point g = (CE
x
,CE
y
);
k is the projection of g on the half-face axis,
with k = (HF
x
,CE
y
) ;
do
line1=pdist2(c,g)
line2=pdist2(g,k)
if NP
y
HF
y
< 0 then
imrotate(180
) Face Flip
end if
if CE
x
HF
x
> 0 then
if (CE
y
== HF
y
) then angle = 90
else angle=radtodeg(sin(line2/line1))
end if
else
if (CE
y
== HF
y
) then angle = 270
else angle=360
-radtodeg(sin(line2/line1))
end if
end if
imrotate(angle) Face Rotation
end do
At first, the training set is made of a set of face
pictures, several poses per person as illustrated in
Fig. 4(a). This training set is subject to facial pre-
processing (Travieso-Gonzalez et al., 2011), which in
this case consists in normalizing the training set by
changing the mean and standard of all the testing im-
ages, in order to reduce errors due to lighting condi-
tion (see Fig. 4(b)).
Hence, let us consider the normalized training set
of face images be Γ
1
, Γ
2
, Γ
3
, ..., Γ
F
. The average face
of the set is defined by:
ψ =
1
F
F
i=1
Γ
i
. (4)
Each face differs from the average by the vec-
tor Φ
i
= Γ
i
ψ. This set of very large vectors de-
fines the face space, and is then subject to Principal
Component Analysis (PCA) for dimension reduction.
Thus, PCA seeks a set of P orthonormal vectors v
i
and
their associated eigenvalues λ
l
which best describes
the distribution data. The vectors v
l
and scalars λ
l
are
the eigenvectors and eigenvalues, respectively, of the
covariance matrix:
ψ =
1
F
F
i=1
Φ
i
Φ
T
i
= AA
T
,
(5)
where the matrix A = [Φ
1
Φ
2
... Φ
F
].
Secondly, in the testing phase, a new face (Γ) is
transformed into its eigenface components, i.e. pro-
jected into the face space, by the operation ω
l
=
v
T
l
(Γ ψ) for l = 1 ... F
0
.
The weights form a vector
T
= [ω
1
ω
2
... ω
F
0
]
that describes the contribution of each eigenface in
representing the input face image, treating the eigen-
faces as a basis set for face images.
To determine which face class provides the best
description of an input face image, one can find the
face class l that minimizes the Euclidean distance
ε
l
= k
l
k, where
l
is a vector describing the
l
th
face class (Turk and Pentland, 1991). Thus, a face
is classified as belonging to class l according to the
nearest neighbour approach, i.e. when the minimum
ε
l
is below a chosen threshold θ
ε
, e.g. as illustrated
for the True Positive (TP) case in Fig. 1 . Otherwise,
the face is classified as ‘unknown’ and an ‘unrecog-
nized pictogram’ is displayed as for the True Negative
(TN) case in Fig. 5.
iFR: Interactively Pose Corrected Face Recognition
109
Figure 5: Example of our iFR system in action, featuring
automatic face detection, interactive pose correction, and
automatic face recognition.
Hence, the automatic recognition of a cropped, de-
tected face is done quickly and precisely as demon-
strated in Section 3.
3 EXPERIMENTS AND
DISCUSSION
To validate our approach, we have applied our iFR
system on the publicly available, well-established
ORL dataset (Gross, 2005).
The ORL database is a set of faces taken between
April 1992 and April 1994 at the Olivetti Research
Laboratory in Cambridge, UK. This dataset contains
400 pictures of 40 distinct subjects. The images are
organised in 40 directories (one for each subject) with
10 images per class. All the images are taken against
a dark homogeneous background and the subjects are
in up-right, frontal position (with tolerance for some
side movements). The files are in portable gray map
(pgm) format, and the size of each image is 92x112,
8-bit grey level.
This database owns challenges such as face varia-
tions over time since some of the pictures were taken
at different times; gender variations since the dataset
contains a mix of images with faces of different men
and women; presence or absence of structural com-
ponents (e.g. beard/no beard, etc.) and variation
of facial details (e.g. glasses/no-glasses, etc.); vari-
ations in facial expressions (e.g. open/closed eyes,
smiling/non-smiling, angry, neutral, etc.); variations
of illuminations conditions; and variations in face
poses such as face in-plane and/or out-plane rotations.
In particular, ORL face poses are within ±20
in yaw
and roll (Zhang and Gao, 2009). All these difficulties
make the dataset challenging and interesting to test
our approach on.
All the experiments have been run on a computer
with Intel Core 2 Duo Pentium T9300, 2.5 GHz, 2Gb
RAM, using MatLab.
We have assessed the performance of our auto-
matic face recognition system by computing the face
recognition rate defined as
recognition accuracy =
T P + T N
T P + T N + FN + FP
, (6)
with TP, true positive, TN, true negative, FP, false pos-
itive, and FN, false negative. The obtained average
face detection rate by our method has been reported
and compared to state-of-art techniques in Table 1.
Experiments have been carried out to test our iFR
system. It consisted in applying the face recognition
method on the ORL database with a training set of 30
pictures (10 subject classes, 3 face poses per class)
and a testing set of 100 pictures (10 subject classes,
10 face poses per class).
Examples of the results of the state-of-art recogni-
tion method (without any pose correction) as well as
the results obtained using our face recognition method
(with pose correction) are presented in Figs. 6. (b)
and (c), respectively.
In Table 1, we have reported the face recognition
scores for the state-of-art eigenfaces which do not in-
clude pose correction, and for our iFR which includes
eigenfaces in conjunction with our Algorithm 1 for
pose correction.
Table 1: Face recognition accuracy scores obtained by
approaches such as (Viola and Jones, 2004), and our
method.
our
recognition accuracy 92% 94%
From Table 1, we can observe that our iFR system
shows excellent accuracy in face recognition, outper-
forming the state-of-the-art eigenface method.
The average computational speed of our approach
is in the range of milliseconds, thus our developed
system could be used in context of online face recog-
nition.
Furthermore, for both speed and accuracy, iFR de-
sign outperforms other approaches using pose correc-
tion such as (Bordoloi and Sarma, 2009), (Nakamura
and Takano, 2011), (Huang et al., 2012).
4 CONCLUSIONS
In this work, we have proposed a computationally ef-
ficient and robust face recognition system called iFR,
consisting of three phases, namely face detection,
BIOSIGNALS 2016 - 9th International Conference on Bio-inspired Systems and Signal Processing
110
(a) (b) (c)
Figure 6: Examples of face recognition results for (a) given
test images; (b) using state-of-art method; (c) our method.
pose correction, and face recognition. Hence, the face
detection phase uses the Viola & Jones method, while
the pose correction phase relies on our new, pose-
correction algorithm fed by interactively selected fa-
cial features, such nose tip and nose top, in the de-
tected face. The face recognition phase applies the
eigenface approach on the pose-corrected face. Our
iFR system combines the advantages of both template
and feature approaches for face recognition. It outper-
forms the state-of-the-art methods, and is well suited
system for online applications involving face recogni-
tion.
REFERENCES
Ahonen, T., Hadid, A., and Pietikainen, M. (2006). Face
description with local binary patterns: Application to
face recognition. IEEE Transactions on Pattern Anal-
ysis and Machine Intelligence, 28(12):2037–2041.
Bordoloi, H. and Sarma, K. K. (2009). Face recognition
with image rotation detection, correction and rein-
forced decision using ANN. International Journal
of Electrical and Electronics Engineering, 3(7):1205–
1212.
Brunelli, R. and Poggio, T. (1993). Face recognition: fea-
tures versus templates. IEEE Transactions on Pat-
tern Analysis and Machine Intelligence, 15(10):1042–
1052.
Choi, J. Y., Neve, W. D., Plataniotis, K. N., and Ro, Y. M.
(2011). Collaborative face recognition for improved
face annotation in personal photo collections shared
on online social networks. IEEE Transactions on Mul-
timedia, 13(1):14–18.
Cootes, T. F., Edwards, G. J., and Taylor, C. J. (2001). Ac-
tive appearance model. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 23(6):681–685.
Du, S. and Ward, R. (2006). Face recognition under
pose variations. Journal of the Franklin Institute,
343(6):596–613.
Gallego-Jutgla, E., de Ipina, K. L., Marti-Puig, P., and Sole-
Casals, J. (2013). Empirical mode decomposition-
based face recognition system. In Proceedings of the
International Conference on Bio-Inspired Systems and
Signal Processing, pages 445–450.
Gross, R. (2005). Face Databases, chapter Handbook of
Face Recognition, pages 301–327. Springer-Verlag.
Han, M.-J., Lin, C.-H., and Song, K.-T. (2013). Robotic
emotional expression generation based on mood tran-
sistion and personality model. IEEE Transactions on
Cybernetics, 43(4):1290–1303.
Huang, G. B., Mattat, M. A., Lee, H., and Learned-Miller,
E. (2012). Learning to align from scratch. In Proceed-
ings of the Conference on Advances in Neural Infor-
mation Processing Systems, pages 764–772.
Jain, A. K., Klare, B., and Park, U. (2012). Face matching
and retrieval in Forensics applications. IEEE Multi-
Media, 19(1):20.
Lindner, C., Bromiley, P. A., Ionita, M. C., and Cootes, T. F.
(2015). Robust and accurate shape model matching
using random forest regression voting. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence,
37(9):1862–1874.
Montagne, C., Kodewitz, A., Vigneron, V., Giraud, V., and
Lelandais, S. (2013). 3D local binary pattern for PET
image classification by SVM. In Proceedings of the
International Conference on Bio-Inspired Systems and
Signal Processing, pages 145–4150.
Murphy-Chutorian, E. and Trivedi, M. M. (2009). Head
pose estimation in computer vision: A survey. IEEE
Transactions on Pattern Analysis and Machine Intel-
ligence, 31(4):607–626.
Nagi, J., Ahmed, S. K., and Nagi, F. (2008). A Matlab-
based face recognition system using image processing
and neural networks. In Proceedings of the Interna-
tional Colloquium on Signal Processing and its Ap-
plications, pages 83–88.
Nakamura, K. and Takano, H. (2011). New Approaces
to Characterization and Recognition of Faces, chap-
ter Face discimination using the orientation and size
recognition characteristics of the spreading associa-
tive neural network, pages 197–212. InTech.
Olszewska, J. I. (2013). Multi-scale, multi-feature vector
flow active contours for automatic multiple-face de-
tection. In Proceedings of the International Confer-
ence on Bio-Inspired Systems and Signal Processing,
pages 429–435.
Rambach, J., Huber, M. F., Balthasar, M. R., and Zoubir,
A. M. (2015). Collaborative multi-camera face recog-
iFR: Interactively Pose Corrected Face Recognition
111
nition and tracking. In Proceedings of the IEEE Inter-
national Conference on Advanced Video and Signal
Based Surveillance, page 1.
Tasli, H. E., den Uyl, T. M., Boujut, H., and Zaharia, T.
(2015). Real-time facial character animation. In Pro-
ceedings of the IEEE International Conference on Au-
tomatic Face and Gesture Recognition, page 1.
Travieso-Gonzalez, C. M., del Pozo-Banos, M., and
Alonso, J. B. (2011). Biometric Systems, Design and
Applications, chapter Facial identification based on
transform domains for images and videos, pages 293–
316. InTech.
Turk, J. and Pentland, A. (1991). Face recognition using
eigenfaces. In Proceedings of the IEEE International
Conference on computer Vision and Pattern Recogni-
tion, pages 586–591.
Viola, P. and Jones, M. J. (2004). Robust real-time face
detection. International Journal of Computer Vision,
57(2):137–154.
Wood, R. and Olszewska, J. I. (2012). Lighting-variable
AdaBoost based-on system for robust face detection.
In Proceedings of the International Conference on
Bio-Inspired Systems and Signal Processing, pages
494–497.
Zhang, X. and Gao, Y. (2009). Face recognition across pose:
A review. Pattern Recognition, 42(11):2876–2896.
BIOSIGNALS 2016 - 9th International Conference on Bio-inspired Systems and Signal Processing
112