Fourier mellin transform based face recognition

INTERNATIONAL JOURNAL OF COMPUTER(IJCET), ISSN IAEME–
International Journal of Computer Engineering and Technology
ENGINEERING
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), ©
0976

& TECHNOLOGY (IJCET)
ISSN 0976 – 6367(Print)
ISSN 0976 – 6375(Online)
Volume 4, Issue 1, January- February (2013), pp. 08-15
IJCET
© IAEME: www.iaeme.com/ijcet.asp
Journal Impact Factor (2012): 3.9580 (Calculated by GISI) ©IAEME
www.jifactor.com

FOURIER MELLIN TRANSFORM BASED FACE RECOGNITION

Sambhunath Biswas1, Amrita Biswas2
System Analyst (GR-I), Machine Intelligence Unit, Indian Statistical Unit, Kolkata, India 1
Associate Professor, Electronics & Communication Engineering, Sikkim Manipal Institute of
Technology, Majitar, India2

ABSTRACT
Human face recognition is, indeed, a challenging task, especially under illumination
and pose variations. We examine in the present paper effectiveness of a simple face
recognition algorithm based on Fourier Mellin Transform. The algorithms convert 2-D gray
level training face images into their respective depth maps or physical shape which are
subsequently transformed by Fourier Mellin Transform. Experiments show that such
transformed shape features are robust to illumination and pose variations. Classification for
test face images is made through a k-NN classifier, based on L1 norm. Proposed algorithm
has been tested on face images from the ORL database.
Keywords: Face Recognition, Depth Map, Fourier Mellin Transform, Nearest Neighbour
Classifier
I. INTRODUCTION

Face Recognition problem has been studied extensively for more than twenty years
but even now the problem is not fully solved. In particular, the problem still exists when
illumination and pose vary significantly. Recently, some progress [1] has been made on the
problems of face recognition, especially under conditions such as smallvariations in lighting
and facial expressions or pose. Of the many algorithms for face recognition, so far developed,
the traditional approaches are based on Principal Component Analysis (PCA). Hyeonjoon
Moon et al. [2] implemented a generic modular PCA algorithm where the numerous design
decisions have been stated explicitly. They experimented with changing the illumination
normalization procedure and studied its effect through the performance of compressing
images with JPEG and wavelet compression algorithms. For this, they varied the number of
eigen vectors in the representation of face images and changed the similarity measure in the
classification process. Kamran Etemad and Rama Chellappa in their discriminant analysis
algorithm [3], made an objective evaluation of the significance ofvisual information in
different parts (features) of a facefor identifying the human subject. LDA of faces provides a
small set of features that carries the most relevant information for classification purposes. The
8

International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME

features are obtained through eigen vector analysis of scatter matrices with the objective of
maximizing between-class variations and minimizing within-class variations. The algorithm
uses a projection based feature extraction procedure and an automatic classification scheme
for face recognition. A slightly different method, called the evolutionary pursuit method, for
face recognition was described by Chengjun Liu and Harry Wechsler [4]. Their method
processes images in a lower dimensional whitened PCA subspace. Directed but random
rotations of the basis vectors in this subspace are searched by Genetic Algorithm, where
evolution is driven by a fitness function defined in terms of performance accuracy and class
separation. Up to now, many face representation approaches have been introduced including
subspace based holistic features and local appearance features [17]. Typical holistic features
include the well-known principal component analysis (PCA) [18], linear discriminant
analysis [19], independent component analysis (ICA)[20] etc. Recently, information from
different sections, such as, scale, space and orientation, has been used for representation and
recognition of human faces by Zhen et al. [21]. This does not include the effect of
illumination change. Subspace based face recognition under the scenarios of misalignments
and/or image occlusions has been published by Shuicheng et al. [22]. We have not considered
image occlusions as our objective is different. The proposed research work addresses the
problem of face recognition to achieve high performance in the face recognition system. Face
Recognition method, [5] based on curvelet based PCA and tested on ORL Database, uses 5
images for training and has achieved 96.6% recognition rate and, using 8 images for training
on the Essex Grimace database, has achieved 100% recognition rate. Another algorithm [6],
based on wavelet transform, uses 5 images for training from the ORL Database has achieved
a recognition rate of 99.5%. But still more improvement is required to ensure that the face
recognition algorithms are robust, in particular to illumination and pose variation. A face
recognition algorithm mainly based on two dimensional graylevel images, in general, exhibits
poor performance when exposed to different lighting conditions. This is because the features
extracted for classification are not illumination invariant. To get rid of the illumination
problem, we have used the 3-dimensional depth images of the corresponding 2-dimensional
gray level face images. This is because the 3-D depth image depicts the physical surface of
the face and thus, provides the shape of human face. The primary reasonis that such a shape
depends on the gradient values of thephysical surface of the face, i.e., on the difference
ofintensity values and not on the absolute values of intensity. As a result, change in
illumination does not affect the feature set and so the decision also remains unaffected. Such
a shape can be obtained using a shape from shading algorithm and subsequently can be used
for feature extraction. 3-D face matching using isogeodesic stripes through a graph as
described in [23] is a different technique for face recognition. But it is computationally
expensive. However, it is also a different area of research. Xiaoyang and Triggs [24], on the
other hand, considered texture features for face recognition under difficult lighting
conditions. Their method needs to enhance local textures but how to select the local textures
or which local textures are adequate and need be considered are not discussed. The proposed
algorithm use the shape from shading algorithm [8], and Fourier Mellin transform
respectively to compute energy for feature extraction. We have used L1 norm distance to test
for classification. With this, the outline of the paper is described as follows: In section II, we
briefly review a shape from shading algorithm and in section III, the concepts of Fourier
Mellin Transform are briefly sketched. Section IV, depicts the proposed algorithm, while
experimental results are discussed in section V. Finally, conclusion is made in the last
section.

9


II. EXTRACTION OF ILLUMINATION INDEPENDENT FEATURES

The problem of recovering 3-D shape from a single monocular 2-D shaded image was
first addressed by B. K. P Horn [14]. He developed a method connecting the surface gradient
(p,q) with the brightness values for Lambertian objects. There result is known as the
reflectance map. Therefore, he computed the surface gradients (p,q) using the reflectance
map in order to get the shape. From (p,q), he also computed depth, Z. Since, orientation of
tangent planes is accompanied by the orientation of their normal vectors, say (nx,ny,nz),they
can also be effectively used to represent the surface shape.As the reflectance map, in general,
is non-linear, it is very difficult to find the gradient values in a straightforward way. Some
other researchers, such as, Bruss [15] and Pentland [16], to simplify the problem, thought of
local analysis to compute the shape. Thus, two different kinds of algorithms, e.g. global and
local emerged. In global methods, Horn showed the shape can be recovered by minimizing
some cost function involving constraints such as smoothness. He used variational calculus
approach to compute the shape in the continuous domain and its iterative discrete version in
the discrete domain. Bruss showed that no shape from shading technique can provide a
unique solution without additional constraint. Later on, P. S. Tsai and M. Shah [8] provided a
simple method to compute shape through linearization of Horn s nonlinear reflectance map.
‟
For our purpose, we have used the shape from shading algorithm described by P. S. Tsai and
M. Shah [8] for its simplicity and fastness. This approach employs discrete approximations
for p and q using finite differences, andlinearizes the reflectance in Z(x,y). The method is
fast, since each operation is purely local. In addition, itgives good results for the spherical
surfaces, unlike other linear methods. Note that the illumination change may be due to the
position change of the source keeping the strength of the source as it is or due to the change
in the source strength keeping the position of the source fixed. In either case, the gradient
values, p and q, of the surface do not change, i.e., they can be uniquely determined [14].
Hence, for the linear reflectance map, the illumination will have no effect on the depth map.
In other words, depth map will be illumination invariant.

III. FEATURE EXTRACTION

Number of methods are available for feature extraction. We have selected Fourier
Mellin Transform based Approach. The Fourier-Mellin transform is a useful mathematical
tool for image recognition because its resulting spectrum is invariant in rotation, translation
and scale. The Fourier Transform itself (FT) is translation invariant and its conversion to log-
polar coordinates converts the scale and rotation differences to vertical and horizontal offsets
that can be measured. A second FFT, called the Mellin transform (MT) gives a transform-
space image that is invariant to translation, rotation and scale.
The Standard Fourier–Mellin Transform is discussed in the following paragraph:
Let f denote a function representing a gray-level image. The standard Fourier–Mellin
transform of fis given by:

(1)
Where Z denotes additive group of integers and R denotes additive group of the real line.
The FMT is a global transform and applies to all pixels the same way. Textured imagescannot
be taken into account directly and objects must first be localized and isolatedfrom the scene

10


to match one of the requirements regarding the existence of the integral inEq. (1).Due to the
singularity at the origin of coordinates, a solution generally adopted is tocancel the image over a small
disk around the origin [9]. However, this approximation has serious effects on the numerical
computation of the FMT because of the following reasons:
(1)Image values nearer the origin have a larger effect on the FMT than image valuesremote from the
centroid because of the 1/rweighting in the measure of the Fourier–Mellinintegrals. Hence, significant
information content of the image is lost in addition to removinga small disk in the image centroid.
(2)It may cause stretching problems when images are enlarged. How large must thedisk be if the
image is scaled by an unknown factor? By cancelling a disk of constant radiusfor every image,
different amounts of information are removed.
More recently, a rigorous approach has been introduced to tackle the difficulties describedabove.
Ghorbel[10] suggested computing the standard FMT offσ(r,θ)=rσf(r,θ) instead of f(r,θ), where σis a
fixed and strictly positive real number.Hence, the integral (1) exists and is called the AFMT of f,with
σ>0

(2)
While the classical Fourier transform converts translation into a pure phase change, the AFMT
converts a similarity transformation in the original domaininto a complex multiplication in the
Fourier–Mellin domain. These relations can be seenas the shift theorem for the planar similarity group
and make the AFMT appropriate forextracting features that are invariant to scale and rotation
changes.[11]
The AFMT can be expressed according to theCartesiancoordinates of f as follows:

(3)
In this case, no resampling of the discrete image is necessary and theAFMTcan be estimateddirectly
from the rectangular grid.TheCartesianAFMT(C-Afmt) approximationis computed by using sums in
place of integrals:

(4)
The coordinates m and n correspond to a pixel position from the object centroid. Pmin, Pmax,
Qminand Qmaxindicate the coordinates, with respect to the image centroid, of the smallest rectangle
that fully contains the object. For the sake of compatibility with otherapproximations, we used the
trapezoidal integration rule. The discrete image is recovered directly in rectangular coordinates.

11


IV. PROPOSED APPROACH

We discuss in this section the proposed approach using Fourier Mellin Transform:
Step1: Compute depths of all the training images using
shape from shading method.
Step2: Compute the Fourier Mellin Transform of the depth image and take the FMT
coefficients as feature vectors.
Step3: Classify the test images using the L1 norm distance measure.
Step4: Stop

V. RESULTS AND DISCUSSION

In order to test the proposed algorithms, we have used ORL. The ORL (AT and T)
database contains 10 different images (92 x 112), each of 40 different subjects. All images
were takenagainst a dark homogenous background with thesubjects in upright, frontal
position with some sidemovement. Sample images of the dataset are shown inFig. 3. The
depth map was computed for all the images in thetraining database assuming the reflectance
of the surface to be Lambertian. The obtained depth image has the same size as the original
image i.e. 92 x112.Depthimage computed by shape from shading algorithm for thefirst image
in ORL database is shown in Fig.1.The Fourier Mellin Transform of the depth map is
computed for feature extraction. The Cartesian approximation of AFMT of the first image of
the ORL database has been shown in Fig.2. To show the robustness of features against
orientation, wehave plottedthe relative error in distance measurement for all tenimages in six
classes (of ORL database) from theirrespective mean images shown in Fig.4. Note thatthis
distance is almost zero for all the images in a class and maintains excellent constancy. We
have tested the algorithm for different number of training images. Classification was
conducted using k-NN classifier based on L1 norm measure The results are shownin Table 1.

TABLE I RESULTS TABLE
Sl.No No.of Training Images Recognition %
1 5 100
2 4 100
3 3 95.7
4 2 90

Fig. 1 Image and its depth map

12


Fig. 2 Illustration of the Cartesian approximation of the AFMT of the original image in Fig.1

Fig. 3 Sample Images of the ORL Database

Fig. 4 Relative error of images from the respective class mean

VI. CONCLUSION

We have proposed a simple algorithm based on image depth map and Fourier Mellin
Transform. The results show that for 4 training images we get 100% recognition percentage
and for 3 training images we get a recognition percentage of 95.7%.This clearly shows that
despite the simplicity of the algorithm we get superior results and there is scope for further
improvement in the recognition percentage by resorting to some superior classification
techniques.
13


REFERENCES

[1] W. Zhao, R. Chellappa, P. J. Phillips, “A. Rosenfeld,FaceRecognition:A Literature
Survey”, ACM Computing Surveys, Vol. 35, No. 4, 2003,pp.399-458.
[2] Hyeonjoon Moon, P Jonathon Phillips, “Computational and Performance Aspects of PCA
Based Face Recognition Algorithms”, Perception 30(3),2001,pp.303 - 321
[3] Kamran Etemad and Rama Chellappa, “Discriminant Analysis for Recognition of Human
Face Images”, Proc. First Int. Conf. on Audio and Video Based Biometric Person
Authentication,Crans-Montana, Switzerland,Lecture Notes In Computer Science; Vol.1206,
August 1997,pp.127 - 142
[4] Chengjun Liu and Harry Wechsler, “Face Recognition using Evolutionary Pursuit”, Proc.
Fifth European Conf. on Computer Vision, ECCV’98,Freiburg, Germany, Vol II, 02-06 June
1998, pp.596-612.
[5] Tanaya Mandal and Q. M. Jonathan Wu, “Face Recognition Using Curvelet” Based PCA,
IEEE, Technical Report, 6/08, pp.978-1-4244-2175.
[6] ZhengDezhong Cui Fayi, ‘Face Recognition based on Wavelet Transform and Image
Comparison”, Proc. International Symposium on Computational Intelligence and Design,
Volume: 2, 2008, pp. 24-29.
[7] C.SydneyBurrus and A. Gopinath and HaitaoGuo, “Introduction to Wavelets and Wavelet
Transforms”, Prentice Hall, N.J 07458, USA, 1998.
[8] Ping-Sing Tsai and Mubarak Shah “Shape From Shading Using Linear Approximation”,
Image and Vision Computing, vol: 12, 1994, pp.487-498.
[9] P. E. Zwicke and Z. Kiss, A new implementation of the Mellin transform and its
application to radar classification, IEEE Trans. Pattern Anal. Mach. Intell. 5, 1983, 191–19
[10] F. Ghorbel, A complete invariant description for gray-level images by the harmonic
analysis approach, PatternRecog. Lett.15, 1994, 1043–1051.
[11] St´ephaneDerrode,Robust and Efficient Fourier–Mellin Transform Approximations for
Gray-Level Image Reconstruction and Complete Invariant escription Computer Vision and
Image Understanding 83, 57–78 (2001)
[12] Peter N. Belhumeur, Joao P. Hespanha and David J. Kriegman,
“EigenfacesvsFisherfaces:Recognition using Class Specific Linear Projection”,IEEE Trans.
on PAMI, July 1997.
[13] R. C. Gonzalez and R. E. woods,” Digital Image Processing”, Dorling Kindersley, India,
Pearson Prentice Hall, 2006.
[14] B.K.P Horn,” Robot Vision”, Cambridge, Massachusetts, USA , MIT Press,1986.
[15] A. R. Bruss, “The Image Irradiance Equation:Its Solution and Applicaion”,Technical
Report TR-623, MIT-AI, June 1981.
[16] A. P. Pentland, “Local Shading Analysis”, IEEE Trans. on PAMI, vol.6,no.2, March
1984, pp.170-187.
[17] S. Z. Li and A. K. Jain, “Handbook of Face Recognition”, New York,Springer-Verlag,
2005.
[18] M. A. Turk and A.P. Pentland, “Face Recognition using eigenfaces”, Proc.IEEE
Computer Society Conf. Comput.vs. Pattern Recognition, June 1991 pp. 586-591.
[19] P. Belhumeur. J. Hespanha and D. Kriegman, “Eigenfaces vs. fisherfaces:recognition
using class specific linear projection”, IEEE Trans. On Pattern Analysis and Machine
Intelligence, vol. 26, no. 9, Sept. 2004, pp.1222-1228.
[20] P. Conor, “Independent component analysis a new concept?”, Signal Processing, vol.
36,1994, pp. 287-314.
14


[21] Zhen Lei, Shengcai Liao, MattiPietikainen and Z. Li “Face recognition by exploring
information jointly in space, scale and orientation”, IEEE Trans. on Pattern Analysis and
Machine Intelligence, vol. 20, no. 1, Jan. 2011, pp.247-256.
[22] Shuicheng Yan, jianzhuang Liu, Xiaoou Tang and Tomas S.Huang,”Misalignment-
robust face recognition”, IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 19,
no. 4, Aril2010, pp. 1087-1096.
[23] Stefano Berretti, Alberto Del Bimbo and Pietro Pala,” 3D face recognition using
isogeodesic stripes”, IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 32, no.
12,Dec. 2010, pp.2162-2177.
[24] Xiaoyang Tan and Bill Triggs,” Enhanced local texture features sets forface recognition
under difficult lighting conditions”, IEEE Trans. on Pattern Analysis and Machine
Intelligence, vol. 19, no. 6, Jun.2010 pp.1635-1650.
[25] Abhishek Choubey and Girish D. Bonde, “Face Recognition Across Pose With
Estimation Of Pose Parameters” International journal of Electronics and Communication
Engineering &Technology (IJECET), Volume3, Issue1, 2012, pp. 311 - 316, Published by
IAEME
[26] Steven Lawrence Fernandes and Dr. G Josemin Bala, “Analysing Recognition Rate Of
Lda And Lpp Based Algorithms For Face Recognition” International journal of Computer
Engineering & Technology (IJCET), Volume3, Issue2, 2012, pp. 115 - 125, Published by
IAEME

15

Fourier mellin transform based face recognition

More Related Content

What's hot (13)

Viewers also liked (7)

Similar to Fourier mellin transform based face recognition (20)

More from iaemedu (20)

Fourier mellin transform based face recognition