Professional Documents
Culture Documents
Research Group in Intelligent Machines (REGIM), University of Sfax, B.P. W, 3038, Sfax, Tunisia
Tel: +21674276400
Fax: +21674274437
Abstract: In recent years, an explosion in research on pattern recognition has been observed.
Face recognition is a specialised pattern recognition task with several applications such as security
(access to restricted areas, banking, identity verification, and people recognition at airports).
Face recognition is an example of advanced object recognition. The process is influenced by several
factors such as shape, reflectance, pose, occlusion and illumination which make it even more
difficult. Today there are many well known techniques for face recognition.
In this paper, we present an approach to face recognition based in Beta Wavelet networks. We then
Introduction:
The Wavelet networks (WNs) theory was introduced in 1992 by Zhang and Benveniste as a
combination of artificial neural radial basis function (RBF) networks and wavelet decomposition.
Since then they have received little attention in recent publications except some searchers who have
used WNs for signal representation and classification. They have explained how a WN can be
-1-
generated and showed how they can be used for pattern matching. In addition, they mention the
large data compression achieved by such a WN representation. Zhang showed that WNs are able to
handle nonlinear regression of moderately large input dimension with sparse training data. Recently
some researchers like kruger and Feris have used WNs to trace and recognize an object using a 2D
functions localized both in time and in frequency. To define these functions, it is necessary to begin
from a function ψ (t ) named ''mother wavelet'' well localized and oscillating (it looks like a wave,
but being localized). The condition of localization is written as the usual shape of decrease in fast
amplitude when t increases indefinitely. The condition of oscillation suggests that ψ (t ) vibrates like
The mother wavelet generates the others wavelets of the family: ψ a,b (t) a>0 and b real, by scaling
(variation of the a parameter) and by shifting in the time (variation of the b parameter).
1 t −b
ψ a,b (t ) = ψ( ) (1)
a a
For the waveletψ a,b , b intervenes in the shifting of the mother wavelet and a in its scaling. This
wavelet is localized in time on the interval [b, b+L*a], if the mother wavelet is localized on [0 , L],
and in frequency on the interval [f0/a , f1/a], if the mother wavelet is localized on [f0 , f1].
According to this definition, we can imagine easily that the high wavelets correspond to a<1
(narrow windows) and the low wavelets correspond to a>1 (large windows).
-2-
Grossman and Morlet showed that when ψ (t ) is a real value this collection can be used as an
orthonormal basis. This means that every finished signal of energy can be written as a linear
combination of waveletsψ a,b (t) and that the coefficients of this combination are the scalar products:
+∞
1 t −b
C ( a, b ) = ∫
a −∞
x(t )ψ (
a
) (2)
warehouse are rarely static, and they have a complex life cycle.
1) Admissibility
Let’s consider a function ψ belonging to L2(R) and TF (ψ ) is its Fourier transform. ψ is called
+∞ TF (ψ (ω )) 2
∫ dω < +∞ (3)
−∞ ω
2) Localization
A wavelet is a function that must have fast decrease on the two sides of its domain of definition or
-3-
3) Oscillation (Moment of order 0)
∫ψ(t)dt =0 (4)
The mother wavelet must satisfy the properties of Shifting and scaling for what can generate other
wavelets.
A wavelet expansion consists in scaling and shifting of only one and same fixed function: the
waveletψ ∈ L2 (ℜ) .
In continuous wavelet transform, the parameters of scaling and shifting vary continuously which
1 t −b
ψ a,b (t) = ψ( ) with a, b ∈ℜ, a ≠ 0 (5)
a a
If the wavelet is admissible, the continuous wavelet transform admits an inverse f(x) as:
-4-
+∞ +∞ w(a, b)ψ ( x − b ) dadb
f ( x) = ∫−∞ (7)
∫−∞
a
The equation (7) gives the expression of an f function with 2 integrations on all possible dilatations
and all possible shifts of the mother wavelet. Suppose that we dispose only a finite number N of
N
f ( x) ≈ ∑ wiψ i (8)
i =1
As an approximation of the relation (7). The finite sum of the relation (8) is an approximation of the
inverse transform. We can view it also as a decomposition of the function f to a sum of coefficients
wi and waveletsψ .
i
The b defines the scaling parameter of the wavelet and the a defines the shifting parameter. Those
parameters may be chosen arbitrarily at this point. According to wavelet theory, any function f
belonging to L2(R) space can be loosely represented by their continuous wavelet transform and
-5-
ψ
w1
x1 ψ w2
∑ yi
....
x2
.
ψ wj
.
....
xd
ψ
In order to find the WN for the function f we minimize, using the LevenbergMarquardt method, the
energy functional:
N
E = min f − ∑ wiψ n (9)
wi ni i =1 i
With respect to the weights wi and the wavelet parameter vectors ni, Equation (8) states that the wi
and ni are optimized, this means that the translation and the dilation of each wavelet are chosen such
so the image is optimally approximated by the weighted sum of wavelets ψni. We therefore define a
• Definition: Let ψni i=1,...,N be a set of wavelets, f a function and wi and ni chosen
according to the energy functional (9). The vector Ψ =(ψ1,...,ψN )T and the matrix w
=(w11,...,wnn ) define then the wavelet network (Ψ,w) for the function f.
-6-
D- Face Recognition by Beta wavelet networks
The rest of this paper will be reserved for the usage of the Beta wavelet networks for face
recognition. This survey will be made through the use of the Beta wavelets to construct a wavelet
networks.
⎧⎛ p q
⎪⎜ x− x0 ⎞⎟ ⎛⎜ x − x ⎞⎟
⎪ 1 si x∈]x ,x [
β(x;p,q, x , x ) = ⎨⎜⎜ x − x ⎟⎟ ⎜⎜ x − x ⎟⎟ 0 1
0 1
⎪⎝ c 0 ⎠ ⎝ 1 c ⎠
⎪ 0 otherwise
(10)
⎩
px + qx
with p,q, x < x ∈ℜ and x = 1 0
0 1 c p+ q
β ( x0 ) = β ( x1 ) = 0
(11)
β ( xc ) = 1
d β ( x ) px1 + qx0 − ( p + q ) x
= β ( x)
dx ( x − x )( x − x )
0 1 (12)
d β ( xc ) d β ( x0 ) d β ( x1 )
= = =0
dx dx dx
p xc − x0
= (13)
q x − xc
1
-7-
d 2 β ( x)
= β ( x) A( x) (14)
dx
1 ⎡ 1 1 ⎤
A(x) = ⎢ − −(p+q)(x+1)+ px + px ⎥ (15)
(x−x )(x −x) ⎢(x −x) (x−x ) 1 0
0 1 ⎣ 1 0 ⎦⎥
We demonstrate in [15] [16] [17] [18] that all the derivatives of the Beta function are wavelets.
We can get different wavelets when we modify the values of the q and p parameters.
Figure : Beta 1 Wavelet (First derivative of the Beta Function) and Beta 2 Wavelet (Second derivative of the Beta Function)
1) Training
The goal was to recognize a face independing of its expression. We proceeded as follows: a WN
(Ψ,w) was optimized for each training face which resulted in a set of template WNs, We chose the
faces with the “normal” expressions to be our gallery faces. For all the training faces we calculate
-8-
the mean square error (MSE) between the original image and the approximated by the network then
1 N M 2
MSE = ∑ ∑ ( A ( i, j ) − B ( i, j ) ) (16)
M * N i =1 j =1
x1 w11 y1
a1
b1
w1j
xj yj
w1m
bk
ak
xm ym
ai
bi
ai: Wavelet scaling parameters wi
bi : Wavelet shifting MSE
parameters
wi: Network connections weight
Base of parameters
(a network / person)
2) Recognition
To recognize a face, we approximate it by the networks then we also calculate the MSE for it. Now
we compare this error to the existing in the base to find the nearest that is generally for the searched
face. The system does not know the face if the error is not near to all the errors of training faces. We
say that an error is not near if its distance passes a fixed threshold.
-9-
Figure 3: Person identified by the system
E- Experimental Results
In this section we will present the results of a face recognition experiment, using the ORL face
The ORL face database contains 40 persons with 10 images per person.
- 10 -
Figure 5:The set of image for one subject of ORL database
The YALE face database has 15 persons with 11 images per person. The individuals show different
- 11 -
Figure 7:The set of image for one subject of Yale database
• First evaluation:
To evaluate our system we calculate the rate of the correct recognitions on a base with 2
persons, we increase after that the number of persons until we evaluate the entire base. We use the
- 12 -
With the ORL Base:
2 20 0 1,000
3 30 0 1,000
4 37 3 0,925
5 45 5 0,900
6 55 5 0,917
7 65 5 0,929
8 75 5 0,938
9 83 7 0,922
10 90 10 0,900
20 168 34 0,832
25 206 44 0,824
30 245 55 0,817
40 365 75 0,830
Table 1: Rate of recognition of ORL faces
80%
75%
70%
65%
60%
2 3 4 5 6 7 8 9 10 20 25 30 40
Nbr of Personnes/base
Recognition rate
- 13 -
With the Yale Base:
2 20 0 1,000
3 29 1 0,970
4 37 3 0,932
5 45 5 0,909
6 54 6 0,909
7 63 7 0,909
8 71 9 0,898
9 76 14 0,859
10 68 18 0,813
11 83 27 0,777
12 90 30 0,773
13 99 31 0,783
14 107 33 0,786
15 113 37 0,776
Table 2: Rate of recognition of YALE faces
100%
90%
Rate
80%
70%
60%
2 3 4 5 6 7 8 9 10 11 12 13 14 15
Nbr of Personnes/base
Recognition rate
- 14 -
• Second evaluation:
To know the minimum number of wavelets necessary to get the maximum number of correct
1 62 10 0,875
2 65 8 0,901
3 66 6 0,925
4 67 5 0,938
10 67 5 0,938
15 67 5 0,938
Table 3: Rate of recognition of ORL faces of 8 persons with different numbers of Wavelets
Recognition rate
100%
95%
Rate
90%
85%
80%
2 3 4 5 6 7
Wavelets number
Recognition rate
Figure 10: Rate of recognition of ORL faces of 8 persons with different numbers of Wavelets
- 15 -
CONCLUSIONS AND FUTURE WORK
In conclusion, a system capable of performing the task of face recognition has been designed. It is
based on a hierarchical wavelet network. Our method consists on two parts: the first is to train
networks on a set of faces images to have a wavelet networks per person, the second is to test it on
The performance of the Beta Wavelet networks as a classifier suitable for face recognition is
To give an increase of performance, all systems with recognition rates in excess of 90% use a
feature extraction stage. It is still believed that with a correct implementation of the feature
extraction stage that the Beta Wavelet networks classifier recognition rate would be in the range of
90 to 95 percent.
- 16 -
REFERENCES
[1] Q. Zhang and A. Benveniste. Wavelet networks. IEEE Trans. Neural Networks, 3:889–898,
1992.
[2] H. Szu, B. Telfer, and S. Kadambe. Neural network adaptive wavelets for signal representation
[3] H. Szu, B. Telfer, and J. Garcia. Wavelet transforms and neural networks for compression and
recognition.
[4] Q. Zhang. Using wavelet network in nonparametric estimation. IEEE Trans. Neural Networks,
[5] C. C. Holmes and B. K. Mallick. Bayesian wavelet networks for nonparametric regression.
[6] L. Reyneri. Unification of neural and wavelet networks and fuzzy fystems. IEEE Trans. Neural
[7] Daubechies. The wavelet transform, time frequency localization and signal analysis. IEEE
[8] R. Feris, V. Krüger, and R. Cesar Jr. Efficient real-time face tracking in wavelet subspace. In
Proceedings of the Int. Workshop on Recognition, Analysis and Tracking of Faces and Gestures
in Real-Time Systems, pages 113–118, Vancouver, BC, Canada, 2001, in conjunction with the
[9] V. Kruger and G. Sommer. Affine real-time face tracking using gabor wavelet networks. In
Proc. Int.Conf. on Pattern Recognition, Barcelona, Spain, Sept. 3-8, 2000. IEEE Computer
Society.
- 17 -
[10] A. Pentland. Looking at people: sensing for ubiquitous and sensable computing. IEEE
[11] V. Kruger, S. Bruns, and G. Sommer. Efficient head pose estimation with gabor wavelet
networks. In Proc. British Machine Vision Conference, Bristol, UK, Sept. 12-14, 2000. BMVC.
[12] J. Bruske and G. Sommer. Dynamic cell structure learns perfectly topology preserving map.
[14] V. Krüger. Gabor wavelet networks for object representation. Technical Report CS-TR-
[15] Chokri BEN AMAR, Mourad ZAIED and Adel ALIMI, Beta Wavelets Synthesis and
[16] Zaied M., Ben Amar C. and Alimi M.A. "Award a New Wavelet Based Beta Function",
International Conference on Signal, System and Design, SSD03, Tunisia, Mars, 2003, vol. 1, P.
185-191.
[17] Bellil W., Ben Amar C. et Alimi M.A. "Beta Wavelet Based Image Compression",
International Conference on Signal, System and Design, SSD03, Tunisia, Mars, 2003, vol. 1,
pp. 77-82.
[18] Bellil W., Ben Amar C., Zaied M. and Alimi M.A. "La fonction Bêta et ses dérivées : vers
une nouvelle famille d’ondelettes", First International Conference on Signal, System and
- 18 -