mapping function of all feed-forward networks. The samples we could acquire the mean vector of each
construction of the RBF neural network involves L
model recorded as θi = (θ i1 ,θ i2 , ,θ ik ) k and
three different layers with feed forward architecture.
Gaussian function is adopted as the membership
variance vector recorded as δ i = (δ i1 ,δ i2 , L ,δ ik )k .
function of neurons in hidden-layer. The output of Where θij δ ij denotes the mean and variance of the
neurons in hidden layer is: jth feature in Pattern i respectively.
x-ci
2
Oik is defined as the current output of the ith
u i = exp(- ) i=1,2,…,n
2σ i 2 neuron in Layer k , uik is defined as the current
Where ci denotes the center of vector for RBF, σ i input of the ith neuron in Layer k.
denotes the radius, u i denotes the output of the ith Layer 1(input layer): the input vector is transfer
to the next layer directly:
neuron in hidden-layer.
The output of neurons in the output layer is: O1i (t) = u1i (t) ˈˈLˈ
i =12 k
m where k represents the dimension of the feature
yk = ∑ wik ⋅ ui i=1,2,…,n; k=1,2,…,m space.
k =1 Layer 2(chaotic and fuzzy layer): In this layer,
Where wik denotes the weight function in the output each node made the input variables chaotic and fuzzy.
layer. The Gauss radial basis function is adopted here as a
fuzzy membership function of neuron. This layer is
4. Structure of chaotic fuzzy RBF neural composed by n neurons, which is divided into m
groups. Each group contains k neurons. Then, n = m
network
h k and input-output relationship of the jth neuron in
the ith group is
Based on the RBF neural network above, we
construct the environment parameters of the network
to give the fuzzy deducing ability to it. The
864
1 [u ij2 (t) − mij (t)]2 1
Oij2 (t) = μ Aij (u ij2 ) = exp{− } The algorithm takes the E (t ) = (T − O 4 ) as the
2πσ ij 2[σ ij (t)] 2 2
Performance indicators to adjust network parameter
in order to minimize the E (t )
Where mij and σ ij represent the mean, and the
According to the negative gradient and chain law,
variance of the Gaussian function separately. μAij the update rule of the adjusting parameters defined as
denotes the degree of membership of the jth character follows
of the input sample attributes to the pattern Ci The w(t + 1) = w(t) + Δw(t) = w(t) + η (−∂E(t) / ∂w(t))
outputs of the neurons in the ith group in the chaotic where W denotes w4 w3 β m γ σ η is
L
fuzzy layer (μ Ai1 ,μAi2 , ,μAik )T is made of the vector learning rate. The gradient of error E with respect to
of membership of the input samples to the pattern W is
Ci .The k-dimensional eigenvector from the input ∂E(t) / ∂w(t) = −e∂O 4 (t) / ∂w(t) ˄˅
layer was translated into the degrees of membership The jth output in layer 4 is
of each character to each pattern by the fuzzification m k
disposal. And that is the fuzzification process. O 4j (t) = ∑ w ij4 (t)∏ (w ij3 / 2πσ ij ) *
i =1 j =1
The Logistic chaotic map is introduced in this
exp{−[ β ij (t ) xi + xij (t ) − mij ] / 2[σ ij (t)] }
c 2 2
layer. First make it digitized:
x(k) = γ (k)x(k-1)(N − x(k-1)) / N ,γ ∈ (3.6,4) From (1)(2)and(3),it is easy to derive the update
The chaos process is adding chaotic control into rule:
input variables, thus the input variable at time t in Δwij4 (t ) = η eOi3 (t )
layer 2 is
Δwij4 (t ) = η ewij4 (t )Oi3 (t ) / wij3 (t )
u ij2 (t) = βij (t)u1i (t) + x ijc (t)
Where Δβ ij (t) = −η ew ij4 (t )Oi3 [ βij (t ) xi + xijc (t ) − mij ]xi /[σ ij (t)]2
x ijc (t) = γ ij (t)x ijc (t-1)(N − x ijc (t-1)) / N
Layer 3(rule layer): Nodes in this layer are called Δmij (t) = η ew ij4 (t )Oi3 [ β ij (t ) xi + xijc (t ) − mij (t )]/[σ ij (t)]2
rule nodes, and each of them denotes a term of rule.
These neurons accomplish map operation from fuzzy Δγ ij (t) = −η ew ij4 (t )Oi3 {[ βij (t ) xi + xijc (t ) − mij ] /[σ ij (t)]2 }*
rule to output space. The output of each node is the
multiplied-weight of all the input signals of the node:
x (t − 1)( N − xijc (t − 1)) / N
c
L, m
k ij
O (t ) = ∏ w O (t)
3
i
3
i
2
ij i = 1, 2, Where e = T − O 4 (t ) T denotes the target output.
j =1
Iterate in this way until the E (t ) became small
DŽ
Layer 4(output layer): The output of this layer is
linear combination of consequences of last layer. enough
That is to say, this layer accomplishment summed-
weight defuses operation. 5. Experiment and discussion
L, m
m
O j (t) = ∑ w (t)u (t)
4 4
ij
4
i j = 1, 2,
i =1
In order to evaluate the proposed face recognition
algorithm, our experiments are performed on two
Where w denotes the weight function in output
4
ij
benchmark face databases: the ORL database and the
layer, u i 4 (t ) = Oi3 (t ) Yale database.
5.1. Testing on the ORL database
4.2. Learning algorithms of chaotic fuzzy
RBF network The original images of the experiment are made
up of 10 faces of person in ORL face database. Each
The following learning algorithm based on of them applies 10 face images with the size of 112
gradient descent. The deviation is the same as that of
the back-propagation (BP) learning law.
h 92 and the gray-level of 256. The chosen of
training set and testing set has two options:
Option 1: The training set is made up of 50 face
images coming from the first 5 ones of every person.
865
While the testing set is formed by 50 face images
coming from the last 5 ones of every person. [2] de Silva L C ,Aizawa K,Hatori M.Detection and
Option 2: The training set is made up of 20 face tracking of facial features by using edge pixel counting and
deformable circular template matching [J].IEICE
images coming from the first 2 ones of every person.
Transactions on Information and Systems, 1995 ,E78-
While the testing set is formed by 80 face images D(9) :1195 - 1207.
coming from the last 8 ones of every person.
ˊ
Table 1 Comparison between CFRBF and other [3] Eleftheriadis A ,Jacauin A. Automatical face location
algorithms (%) detection for model-assisted rate control in H.261-
compatible coding of video[J]. Signal Processing:Image
Communication ,1995,7(4-6):435-455.
The comparison of CFRBF to the nearest [10] Cox T, Cox M. Multidimensional Scaling1London:
neighbor is shown in Table 2. It can be seen that the Chapman & Hall, 1994
classifier based on CFRBF gains a better
generalization capability and has unique advantages [11] S. Roweis , L. Saul . Nonlinear dimensionality
in dealing with few samples. Table 3denotes different reduction by locally lin2ear embedding[J].
Science ,2000 ,290 (5500) : 2323 - 2326.
performance when the methods extracting the feature
is different. It is shown that feature extraction can [12] J . B. Tenenbaum,V.de Silva,J.C.Langford.A global
improve the recognition accuracy effectively. The geometric framework for nonlinear dimensionality
combination between CFRBF and feature extraction reduction[J].Science,2000, 290(5500) :2319 - 2322.
possesses of a prospective potential.
ˊ
Table 3 The recognition accuracies of different [13] Wang Qing-chun,Wang Hui. Application of Isomap
feature extractions and RBF Neural Network in Face Recognition System [J].
Computer Development & Applications,19(10)
6. References
[1] Ballard P,Stockman G C.Controlling a computer via
facial aspect[J]. IEEE Transaction on System,Man and
Cybernetics , 1995 ,25 (4) :669 - 677.
866