MONOGENIC SCALE SPACE BASED REGION COVARIANCE MATRIX DESCRIPTOR FOR FACE RECOGNITION

David C. Wyld, et al. (Eds): CCSEA, SEA, CLOUD, DKMP, CS & IT 05, pp. 169–177, 2012.
© CS & IT-CSCP 2012 DOI : 10.5121/csit.2012.2218
MONOGENIC SCALE SPACE BASED REGION
COVARIANCE MATRIX DESCRIPTOR FOR FACE
RECOGNITION
M. Sharmila Kumari*
, B H Shekar#
and G Thippeswamy$
*
Department of Computer Science and Engineering, P A College of
Engineering, Mangalore, Karnataka, India.
#
Department of Computer Science, Mangalore University, Karnataka, India.
$
Department of Computer Science and Engineering, Sir MVIT, Yalahanka,
Bangalore, Karnataka, India.
{sharmilabp, bhshekar, swamy.gangappa}@gmail.com
ABSTRACT
In this paper, we have presented a new face recognition algorithm based on region covariance
matrix (RCM) descriptor computed in monogenic scale space. In the proposed model, energy
information obtained using monogenic filter is used to represent a pixel at different scales to
form region covariance matrix descriptor for each face image during training phase. An eigen-
value based distance measure is used to compute the similarity between face images. Extensive
experimentation on AT&T and YALE face database has been conducted to reveal the
performance of the monogenic scale space based region covariance matrix method and
comparative analysis is made with the basic RCM method and Gabor based region covariance
matrix method to exhibit the superiority of the proposed technique.
KEYWORDS
log-Gabor Transform, Riesz transform, Monogenic scale space, Region covariance matrix
descriptor, Face recognition.
1. INTRODUCTION
The biometrics is the well addressed research field in the domain of computer vision and we have
seen plethora of algorithms in particular on face recognition because of its wide acceptability in
several applications ranging from identity authentication, access control and face based video
indexing/retrieval to human–computer interaction/communication. In the recent days, among the
several models proposed for face recognition, descriptor based algorithms gain much importance
because of their robustness against illumination, noise, and occlusion – the common problems
encountered in real recognition environment.
The local descriptors [8, 14, 22] are commonly employed in a number of real-world applications
such as object recognition [5, 14] and image retrieval [16] as they can be computed efficiently,
are resistant to partial occlusion, and are relatively insensitive to changes in viewpoint.
Mikolajczyk and Schmid [15] presented a comparative study of several local descriptors
including steerable filters [7], differential invariants [10], moment invariants [22], Scale Invariant

170 Computer Science & Information Technology (CS & IT)
Feature Transform [13], and cross-correlation of different types of interest points [8, 16]. Their
experiments showed that the ranking of accuracy for the different algorithms was relatively
insensitive to the method employed to find interest points in the image but was dependent on the
representation used to model the image patch around the interest point. In the SIFT algorithm,
each keypoint is represented by its neighbourhood, described as a set of orientation histograms
computed from the gradient image. The SIFT descriptors are invariant to scale, rotation, lighting
and viewpoint change. The most common implementation uses a 128 dimensional descriptor. Ke
and Sukthankar [9] proposed PCA-SIFT descriptor which is also based on the gradient image, the
main difference with SIFT being the further compression using PCA. Recently, Speeded Up
Robust Features (SURF) descriptor [1] has appeared as an alternative to SIFT. Its main advantage
is its fastest computation, while keeping a high descriptive power. It is partially inspired by SIFT,
but instead of using the gradient image, it computes first order Haar wavelet responses. The
LESH or Local Energy based Shape Histogram has been specifically designed for face
recognition applications. Its goal is to encode the underlying shape present in the image.
Basically, the descriptor is a concatenation of histograms obtained by accumulating local energy
along several filter orientations.
On the other hand, we have seen appearance based approaches which are capable of withstanding
noise and occlusion and simple in terms of implementation. In these approaches, data
transformation is a fundamental step and the goal is to obtain highly discriminative lower-
dimensional data from high-dimensional data. Principal component analysis (PCA) and linear
discriminant analysis (LDA) are the widely used tools in the face recognition domain, which
encode high-dimensional face images as lower-dimensional eigenfaces [20] and fisherfaces [2]
respectively. PCA is a linear method that ensures that the data transformed are uncorrelated and
preserve maximally the second order statistics of the original data, and hence is insensitive to the
dependencies of multiple features in the patterns. To overcome this problem, kernel PCA [19] is
proposed as a non-linear extension of PCA that computes the principal components in a high-
dimensional feature space. On the similar line, kernel FLD is proposed as a non-linear extension
to linear discriminant analysis.
We have also seen the frequency domain based approaches where high-frequency components are
used as facial features because of their robustness to illumination changes. The local phase
information is used in some techniques as it is proved to be sufficient to completely reconstruct a
signal within a scale factor [17]. As the Gabor wavelet representation captures salient visual
properties such as spatial localization, orientation selectivity, Lades et al. [11] applied Gabor
wavelets for face recognition using dynamic link architecture framework. Wiskott et al. [23]
extended this basic framework to devise Gabor wavelet based elastic bunch graph matching
method to label and recognize faces. Liu [12] proposed Gabor-based kernel PCA method by
integrating the Gabor wavelet representation of face images and the kernel PCA method for face
recognition. Porikli and Tuzel [18] and Tuzel et al. [21] proposed a new descriptor framework
called region covariance matrices (RCM) for object detection and tracking. The RCMs can be
categorized as a matrix-form feature. However, direct application of RCM to face recognition has
not produced satisfactory results and hence Yanwei et al. [24] introduced Gabor-based region
covariance matrices as face descriptors. The results of Yanwei et al. [24] are highly encouraging
when compared to any other Gabor wavelet based techniques. Although Gabor wavelet based
face recognition models possess very high recognition accuracy irrespective of noise, occlusion
and illumination problems, massive computing and space requirements are the major bottleneck
in these approaches. In this context, we have proposed a new frame work for face recognition
based on monogenic scale space model. The monogenic signal requires fewer convolutions when
compared to Gabor wavelet based model and hence require much lesser computing time. The
region covariance matrix descriptor is employed on monogenic scale space based images to
obtain the representation of face images.

Computer Science & Information Technology (CS & IT) 171
The remaining part of the paper is organized as follows. In section 2, we review monogenic scale
space technique. The region covariance matrix descriptor and its construction in monogenic scale
space for face recognition is presented in sections 3 and 4. The experimental results and
comparative study is given in section 5. Conclusion is reached in section 6.
2. MONOGENIC SCALE SPACE: A REVIEW
The monogenic signal [3, 4] is based on the Riesz transform which is used instead of the Hilbert
transform. The monogenic signal analysis [3, 4] is a framework to interpret images in terms of the
local phase, local orientation and local energy. The monogenic signal is an effective tool to
analyze 2-D signals in a rotation invariant manner. The signal is built upon the first order Riesz
transform. The spatial representation of the Riesz kernel in 2D space is:
൫ܴ௫ሺ‫ܠ‬ሻ, ܴ௬ሺ‫ܠ‬ሻ൯ = ൬
‫ݔ‬
2ߨ|‫|ܠ‬ଷ
,
‫ݕ‬
2ߨ|‫|ܠ‬ଷ൰ , ‫ܠ‬ = ሺ‫,ݔ‬ ‫ݕ‬ሻ߳ℝଶ
(1)
and its transfer function in the Fourier domain is:
ሺ‫ܨ‬௨ሺ‫ܝ‬ሻ, ‫ܨ‬௩ሺ‫ܝ‬ሻሻ = ൬−݅
‫ݑ‬
|‫|ܝ‬
, −݅
‫ݒ‬
|‫|ܝ‬
൰ , ‫ܝ‬ = ሺ‫,ݑ‬ ‫ݒ‬ሻ߳ℝଶ
(2)
For any image, say I(x), the monogenic signal is defined as the combination of I and its Riesz
transform:
i.e., Im(x)= (I(x), Rx{I}(x), Ry{I}(x))
= (I, Rx * I, Ry * I)
(3)
where * stands for the convolution operation.
And hence the local orientation is calculated as [3, 4]:
ߠ = arctan
ܴ௬{۷}
ܴ௫{۷}
, ߠ ߳ [0, ߨሻ (4)
The local phase is defined as:
߮ = ܽ‫2݊ܽݐ‬ ቆටܴ௫
ଶ{۷} + ܴ௬
ଶ{۷} , ۷ ቇ , ߮߳[0, ߨሻ (5)
The local energy is defined as:
‫ܧ‬ = ቆටܴଶ{۷} + ܴ௫
ଶ{۷} + ܴ௬
ଶ{۷} , ۷ ቇ (6)
where R2
{I} = I * F-|
(G(w)).Here G(w) is the log-Gabor filter in the Fourier domain.

Since log-Gabor filters are band-pass filters, usually multi-scale monogenic representation is
required to fully describe a signal. In figure 1, we have given the convolved images in monogenic
scale space filter showing the log-Gabor transformed image with its energy and orientation
images. One can see that the local structure is well captured in monogenic components.
Fig. 1. (a) Original image; (b) log-Gabor image; (c) Energy image and (d) Orientation image
3. REGION COVARIANCE MATRIX: A REGION STATISTICS BASED DESCRIPTOR
The region covariance matrix (RCM) proposed by Tuzel et al. [21] is an efficient matrix form
descriptor which is defined as the covariance of many image statistics computed inside the region
of an image. Let I be an image of size MxN. Define a function ɸ that extracts d dimensional
feature vector zi from a pixel at (x, y) of I. i.e.,
ɸ(I, x, y) = zi Є Rd
(7)
where i = y x M + x is the index of (x, y). Considering all the pixels in a region R, it is
represented as dxd covariance matrix (CR) of the feature points zi.
i. e. , Cୖ =
1
n − 1
෍ሺz୧ − μୖሻሺz୧ − μୖሻ୘
୬
୧ୀଵ
(8)
where µୖ is the mean of z୧
In [21], the function ɸ is defined by the pixel location, intensity value, (colour values (RGB) in
case of colour image) and the norm of the first and second order derivatives of the image.
i.e., ɸ(I, x,y) = [x y I(x,y) |Ix| |Iy| |Ixx| |Iyy|], (9)
where |.| denotes the absolute operator, Ix and Iy respectively denote the first order partial
derivative wrt x and y and, Ixx and Iyy denote the second order partial derivative wrt x and y. One
can notice here that the RCM is a symmetric matrix and capable of capturing both spatial and
statistical properties as diagonal entries represent the variance of each feature and non-diagonal
entries represent their respective correlations.
For classification purpose, the distance measure proposed by Forstner and Moonen [6] is used
which is an eigen-value based distance measure.

∂ሺCRଵ, CRଶሻ = ඩ෍ lnଶ൫λ୧ሺCRଵ, CRଶሻ൯
ୢ
୧ୀଵ
(10)
where λ1,λ2,…λd are the generalized eigenvalues of CR1 and CR2 computed from:
λiCR1ui = CR2ui, i=1,2,…,d.
Although RCM based methods are successful in people detection, object tracking, texture
classification etc., they exhibit poor performance for face recognition problems. Hence, we
proposed to extract monogenic signal based RCM, which possess better recognition rate when
compared to the state of the art face recognition algorithms.
4. CONSTRUCTION OF REGION COVARIANCE MATRIX DESCRIPTOR IN
MONOGENIC SCALE SPACE.
The monogenic scale space based features are obtained by employing Riesz transform. The
monogenic filter is employed on an intensity image, say I at different scales to obtain the local
energy, local phase and local orientation information. In our work, we have considered local
energy at five different scales for constructing RCM which is as follows.
It shall be observed from Eq. (6) that the energy can be computed at different scales. Let A1,…,
A5 be the energy computed at five different scales. Hence, the feature mapping for each pixel (x,
y) is defined as:
ɸ(I, x,y) = [A1(x, y) A2(x, y) A3(x,y) A4(x,y) A5(x,y)]
It shall be observed here that the size of the RCM is 5x5 = 25. One can notice here that the
proposed model requires only five convolutions with the Riesz transform where as Gabor wavelet
based representation requires 40 convolutions (5 orientation and 8 scales) with Gabor filter and
hence the feature mapping results in 40x40 dimension feature vector which is very massive and
hence consume much memory to store the feature vector. Obviously the computing time increases
due to more number of faces in image databases.
So far, we have considered whole image as an input. It is possible to take upper half, lower half,
left half or right half of an image and could construct RCM possibly to handle occlusion.
As suggested in [19], the eigen-value based distance measure (Eq. (10)) is considered for finding
the similarity between two faces images.
5. EXPERIMENTAL RESULTS
This section presents the results of the experiments conducted to corroborate the success of the
proposed model. We have conducted experimentation using two benchmark face image databases
namely AT&T and YALE face databases. All the experiments are conducted using P-IV machine
on Windows-7 platform using MATLAB 7.8 tool.
Experimentation on AT&T face databases: The AT&T face database consisting of 400 gray-
scale images of 10 subjects, each covering a wide range of poses as well as race, gender and

appearance. In Fig. 2, we have shown subset of one such subject of the AT&T database. The
experimentation consists of varying number of training samples and testing samples under each
class. We have chosen first five samples for training and the remaining samples for testing for
each person and the recognition accuracy is obtained. Similarly, two other test cases are generated
which consists of first four and three samples for training and the remaining samples for testing.
The results obtained due to the proposed model are reported in Fig. 3. The results obtained due to
Gabor wavelet based RCM and basic RCM are also shown in Fig. 3. It shall be observed that the
recognition accuracy of the proposed model is on par with the Gabor wavelet based RCM, and
quite high when compared to the basic RCM approach. Over all, the performance of the proposed
model is good as it consumes less time with good recognition results.
Fig. 2. Subset of one subject of AT&T face database
Fig. 3. Recognition accuracy of the proposed model and other techniques on AT&T face
database.
Experimentation on YALE face databases: The YALE face database contains 165 images of
15 subjects that include variation in both facial expression and lighting. In Fig. 4, we have shown
the closely cropped images which include internal facial structures. Here, we have made the
following type of testing. The training set comprised of six images randomly chosen for each
person with remaining number of face images for each person. The recognition accuracy is
computed for the proposed model, Gabor wavelet based RCM technique and the basic RCM

technique. Similarly, four and five face images are chosen randomly under each person during
training and the remaining face images are used for testing. The results are shown in Fig. 5. It
shall be observed from Fig. 5 that the proposed model posses better performance.
Fig. 4. Subset of one subject of YALE face database
Fig. 5. Recognition accuracy of the proposed model and other techniques on YALE face
database.
6. CONCLUSION
The monogenic scale space based region covariance matrix is developed for face recognition. The
energy information contained in the monogenic signal at different scales is used to build the
region covariance matrix descriptor for each face image. Experimental results on the standard
benchmark databases reveal the superiority of the proposed model for face recognition problems
and its suitability in real environment for face recognition problems.

REFERENCES
1. Bay H, Ess A, Tuytelaars T, Gool L V, (2008). “Speeded-Up Robust Features (SURF)”, Computer
Vision and Image Understanding, Vol. 110, Issue 3, pp. 346-359.
2. Belhumeur, P.N., J.P. Hespanha, and D.J. Kreigman, (1997). “Eigenfaces vs. Fisherfaces:
Recognition using class specific linear projection”, IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 19(7), pp. 711-720.
3. Felsberg, M. and Sommer, G. (2004). “The monogenic scale-space: A unifying approach to phase-
based image processing in scalespace”, Journal of Mathematical Imaging and Vision, 21(1), pp.5–26.
4. Felsberg, M., and G. Sommer, (2001). “The monogenic signal”, IEEE Transactions on Signal
Processing, vol. 49, pp. 3136-3144.
5. Fergus, R., P. Perona, and A. Zisserman, (2003). “Object class recognition by unsupervised scale-
invariant learning”, In Proceedings of Computer Vision and Pattern Recognition, pp. 264-271.
6. Forstner, W. and Moonen, B, (1999). “A metric for covariance matrices”, Technical report, Dept. of
Geodesy and Geoinformatics, Stuttgart University.
7. Freeman, W.T. and E. H. Adelson, (1991). “The design and use of steerable filters”, IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 13(9), pp. 891–906.
8. Harris, C. and M. Stephens, (1988). “A combined corner and edge detector”, In Alvey Vision
Conference, pages 147–151.
9. Ke, Y. and R. Sukthankar, (2004). “PCA-SIFT: A more distinctive representation for local image
descriptors”, In Proceedings of IEEE International Conference on Computer Vision and Pattern
Recognition, Washington, pp. 506-513.
10. Koenderink, J. and A. van Doorn, (1987). “Representation of local geometry in the visual system”, In
Biological Cybernetics, Vol. 55, pp. 367–375.
11. Lades, M., J.C. Vorbruggen, J. Buhmann, J. Lange, C. von der Malsburg, R.P. Wurtz, and W. Konen,
(1993). “Distortion Invariant Object Recognition in the Dynamic Link Architecture”, IEEE
Transactions on Computers, vol. 42, pp. 300-311.
12. Liu, C., (2004). “Gabor-Based Kernel PCA with Fractional Power Polynomial Models for Face
Recognition”, IEEE Transactions on Pattern Analysis And Machine Intelligence, Vol. 26 (5), pp. 572-
581.
13. Lowe, D.G., (2004). “Distinctive image features from scale-invariant keypoints”, International
Journal of Computer Vision, vol. 60(2), pp. 91-110.
14. Lowe, D.G., (1999). “Object recognition from local scale-invariant features”, In Proceedings of
International Conference on Computer Vision, pp. 1150–1157.
15. Mikolajczyk, K and C. Schmid, (2005). “A performance evaluation of local descriptors”, IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 27(10), pp. 1615-1630.
16. Mikolajczyk, K and C. Schmid, (2001). “Indexing based on scale invariant interest points”. In
Proceedings of International Conference on Computer Vision, pages 525–531.
17. Oppenheim, A.V. and 1. S. Lim, (1981). “The importance of phase in signals”, Proceedings of The
IEEE, vol. 69, no. 5.
18. Poriklim F and O. Tuzel, (2006). “Fast construction of covariance matrices for arbitrary size image
windows”, In Proceedings of IEEE International Conference on Image Processing, pp. 1581-1584.
19. Schokopf, B., S. Mika, C.J.C Burges, P. Knirsch, K.-R Muller, G. Ratsch, and A.J. Mola, (1999).
“Input space versus feature space in kernel based methods”, IEEE Transactions on Neural Networks,
vol. 10(5), pp. 1299-1319.
20. Turk, M. and A. Pentland, (1991). “Eigenfaces for recognition”, Journal of Cognitive Neuroscience,
vol. 3(1), pp. 71-86.
21. Tuzel, O., F. Porikli, and P. Meer, (2006). “Region covariance: a fast descriptor for detection and
classification”, In the Proceedings of European Conference on Computer Vision, pp. 589-600.
22. Van Gool, L., T. Moons, and D. Ungureanu, (1996). “Affine/photometric invariants for planar
intensity patterns”, In Proceedings of European Conference on Computer Vision. pp. 642-651.
23. Wiskott, L., J.M. Fellous, N. Kruger, and C. von der Malsburg, (1997). “Face Recognition by Elastic
Bunch Graph Matching,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 19,
no. 7, pp. 775-779.
24. Yanwei P., Yuan Yuan and Xuelong Li, (2008). “Gabor-based region covariance matrices for face
recognition”, IEEE Transactions on circuits and systems for video technology, vol. 18(7), pp. 989-
993.

AUTHORS
M. Sharmila Kumari obtained her B.E., degree from University of Mysore in the year 2000 and M.Tech.,
degree in Computer Science and Technology from Visweswaraya Technological University in the year
2004. She is currently working as an Associate Professor in the Department of Computer Science and
Engineering, PA College of Engineering, Mangalore, India. She has authored about twenty five peer-
reviewed papers in International Journals and Conferences. Her areas of research cover Video Image
Processing, Biometrics and Content Based Image Retrieval.
B. H. Shekar obtained his B.Sc., M.Sc., and Ph.D., degrees in Computer Science and Technology from
the University of Mysore, Mysore, India, respectively in the years 1992, 1994, and 2007. He is currently
working as an Associate Professor in the Department of Studies in Computer Science, Mangalore
University, Mangalore. He has authored about 15 Journal papers and 50 peer-reviewed International
Conference papers. His areas of research cover Video Image Processing, Biometrics, Object recognition
and Shape analysis.
G. Thippeswamy obtained his B.E., and M.E., degrees in Computer Science and Technology from
Visweswaraya Technological University. He is currently working as an Assistant Professor in the
Department of Computer Science and Engineering, Sir M Visvesvaraya Institute of Technology,
Yalahanka, Bangalore, India. He has authored about fifteen peer-reviewed papers in International Journals
and Conferences. His areas of research cover Biometrics and Content Based Image Retrieval.

MONOGENIC SCALE SPACE BASED REGION COVARIANCE MATRIX DESCRIPTOR FOR FACE RECOGNITION

More Related Content

What's hot (20)

Similar to MONOGENIC SCALE SPACE BASED REGION COVARIANCE MATRIX DESCRIPTOR FOR FACE RECOGNITION (20)

More from cscpconf (20)

Recently uploaded (20)

MONOGENIC SCALE SPACE BASED REGION COVARIANCE MATRIX DESCRIPTOR FOR FACE RECOGNITION