Face Recognition Using Sign Only Correlation

Document Sample

```					                                       ACEEE International Journal on Signal and Image Processing Vol 1, No. 1, Jan 2010

Face Recognition Using Sign Only Correlation
Ganesh V. Bhat1, K. K. Achary2
1
Department Of Electronics, Canara Engineering College, Bantwal, India
Email: ganeshvbhat@yahoo.com
2
Department Of Statistics, Mangalore University, Mangalore, India
Email: kka@mangaloreuniversity.ac.in

Abstract—In this work we examine a face recognition system
For a given set of vectors X = [ xi ,..., x M ] the
based on advanced correlation filters. A thorough
theoretical design analysis of, Minimum Average                       correlation function between a vector x i and a filter h is
Correlation Energy (MACE) filters and Optimum Trade-off
defined as
Synthetic Discriminant Function (OTSDF) also commonly
known as Optimal Trade-off Filter (OTF) is provided. In                                                  D
practice one of the major computational aspects in the                            σ i (t ) = xi ∗ h = ∑ ( xij ) h j+t                (1)
correlation filter design is representation of complex                                                   j =1
floating point Discrete Fourier Transform (DFT)
coefficients using limited precision memory. In order to over             In SDF the values of the different correlation function
come the floating point memory requirement of the                     at the origin is constrained to some preset value, i.e.
correlation based filters for systems with limited
computational resources use of Discrete Cosine Transform                                   D
Sign-Only Correlation (DCTSOC) which deals with only the                   vi = σ i (0) = ∑ xij h j = xT h
i
i = 1, ..., M   (2)
has been proposed. The proposed method is tested for
Equation 2 can be represented in matrix form as,
synthesis of OTF and a comparison of recognition rate for
frontal face identification is made between OTF using
X h=v
T
DSOC (OTDS) and standard OTF                                                                                                         (3)
Index Terms—correlation filters, discrete cosine transform,           where, v is the desired output response of the filter h . In
face recognition, synthetic discriminant function                     practice it turns out that for cases where M < D the
desired system has infinitely many solutions. If a unique
I. INTRODUCTION                                  solution in column space X is chosen to be h = X ϑ for a
Correlation filters have been applied successfully for             vector ϑ , the filter response can be expressed as
variety of applications, such as object tracking in real                                                 T      −1
time[1], automatic target recognition [2] and recognition                                      h = X(X X) v                          (4)
of biometrics e.g., face, finger print and iris [3]. It is well
where v i ’s are generally chosen to be one (output
known that matched spatial filters are optimal in terms of
maximum output signal to noise ratio for the detection of             constrain), the aim of the constraints is to control the
known image in the presence of noise under the                        variation in the output due to changes in rotation and
assumption of Gaussian statistics [1]. One of the most                scale. Under the assumption that columns of X are
well known of such composite linear correlation filters is            independent, the SDF approach as given in equation 4
the OTF’s [4], which have been proposed as to improve                 attempts to design a filter robust against scale and
the generalizing properties of the MACE filters [5]. In               rotation by constraining the output to a specified value.
this paper we propose a new version of OTF based on                   The issue of invariance to distortion or noise is addressed
DCT sign only correlation termed as OTDS which aims                   by a variant of SDF called as Minimum Variance
to simplify the image representation and correlation                  Synthetic Discrimenant Function (MVSDF). In MVSDF
aspects of the OTF filter design. In section 2, a thorough            the correlation filter is designed assuming the presence of
theoretical design analysis of, MACE filter and OTF is                additive stationary noise denoted by n i.e.
provided. In section 3 and 4 we formulate DCT sign only
correlation as an alternate to DFT based correlation                                             x i = x i +n
%                             (5)
approach used in the design of standard OTF. In section
6, we present simulation results of frontal face                      where x i represents the noiseless image, under the
%
recognition systems designed using OTDS approach and
standard OTF approach.                                                assumption that the noise process is characterized by zero
mean and covariance matrix C . The output of the filter at
II. CORRELATION FILTERS BACKGROUND                           the origin can be written as,
Since the first proposal of Synthetic Discriminant
Function (SDF) [6] there has been a great deal of interest                          σ i (0) = h T x i = h T x i + h T n
%                        (6)
in the concept of SDF and its variants.

1
DOI: 01.ijsip.01.01.01
ACEEE International Journal on Signal and Image Processing Vol 1, No. 1, Jan 2010

The variance in the output denoted by σ (0) is given                                                          x i along the diagonal                     and F ( xi ) represents the
by ρ = h Ch . Minimizing this output variance reduces
2        T
Discrete Fourier Transform of x i . Thus the energy in
the impact of n . The solution of h which minimizes ρ
2
Fourier domain can be expressed as,
under the constrain given by equation 3 is obtained using
Ei = F (h ) diag (F ( xi )) diag ((F ( xi )) F (h )
H                                             H
(13)
the concept of Lagrange multipliers
H
where A represents the hermitian of A . Minimizing
T                        T            T
f ( h , λ ) = h Ch − λ ( X h − v ) =                                                   the average energy by summing Ei ’s over i and dividing
by M the average correlation energy can be written as
D        D                       D
⎛        D
⎞                  M
Eavg = F (h ) BF (h ) ,where
2∑ ∑ Cij hi h j + ∑ Cij ( hi ) − ∑ λi ⎜ −vi − ∑ xij h j ⎟ (7)
H
2

i =1 j = i +1    i =1           i =1 ⎝       j =1      ⎠
⎛ 1
∑ ( F ( xi ) F ( xi ) ) ⎞
M
B = diag ⎜                                  ⎟
*
Since C is symmetric                                                                                                                                                                        (14)
⎝M    i =0                    ⎠
δf            D                            M

= 2∑ Cin hi − ∑ λi xin = 2Ch - Xλ = 0                                                (8)           By analyzing Fourier transform in terms of matrix
δ hn              i =1                         i =1
multiplication we know that       F = (1 / D) F
−1                 H
hence
Equation 8 represents a system of D equations, the                                                           equation 1 can be expressed in Fourier domain as
solution for which is obtained as,
F ( X ) F (h ) = Dv
H
(15)
⎛1 ⎞
h = C X⎜ λ ⎟
-1
(9)           Applying the same analysis as used for MVSDF in
⎝2 ⎠                                                                     Fourier domain we have the final equation for MACE
filter given as
Where λ = [ λ1                    λ2                ...             λM ] , substituting equation
(                              )
−1                      −1                −1
F (h ) = B F ( X ) F ( X ) B F ( X )
H
9 in equation 3 we get                                                                                                                                                              Dv        (16)

T
X h = X C X⎜
T       -1           ⎛ 1 λ⎞ = v ,                            Although the MACE filter does ensure maximum at
⎟
⎝2 ⎠                                 the origin, its ability not to generalize for image
characteristics unseen during training makes it
particularly successful for verification and registration.
⎛ 1 λ⎞
( X C X)
T                        −1
v=⎜            ⎟
-1
(10)        One of the drawbacks of MACE filter is that, in order to
⎝2 ⎠                                produce very sharp peaks in the correlation plane for
authentic images the MACE filter emphasizes on high
Using equation 9 and equation 10 we get the solution
spatial frequencies which makes MACE filters
for MVSDF as
susceptible to noise.

(                            )
-1                   T        -1           −1                              Optimal Trade-off Filter which is a variant of the
h=C X X C X                                                    v           (11)        MACE filter minimizes the average correlation energy
with the addition of an appropriate noise factor, typically
One of the major drawbacks with classical SDF and                                                           modeled as Gaussian with diagonal covariance matrix
MVSDF is that they only constraint the output at a single                                                        C .Under the assumption that the effect of the noise on
point at the origin of the correlation plane without
the Fourier transform of the image is independent and
maximizing it, MVSDF in addition introduces the matrix
uncorrelated with the images in the training set, the
C of size D × D which may be difficult to estimate and
variation of the noise in the correlation output can be
can increase the computational burden. The MACE filter                                                                                                 H
aims to minimize the correlation energy over all the                                                            shown to be given by F (h ) CF (h ) [7], where C is a
images in the training set which results suppression of all                                                     diagonal matrix of dimension D × D containing input
values in the correlation plane except at the origin,                                                           noise power spectral density values as its diagonal
making the origin attain the maximum value in the                                                               elements. When input noise is modeled as white, i.e.
correlation plane.                                                                                               C = I minimizing output variance F (h ) CF (h ) results
H

Making use of the frequency domain properties of
in a filter that emphasizes low spatial frequencies, where
correlation, the correlation between the input image
as minimizing the average correlation energy
x i and the filter h in frequency domain is given as,                                                               H
F (h ) BF (h ) leads to a filter that emphasizes high spatial
∗
F ( xi ) F (h ) = diag (F ( xi )) F (h )
∗
(12)        frequencies, hence an optimal trade off between
minimizing output variance and minimizing average
∗                                                                                          correlation energy results in OTF [8] which minimizes
Where a represents the complex conjugate of a ,
F (h ) TF (h ) subject to the constraints in equation 3,
H

diag ( x i ) represents diagonal matrix with the element of

2
DOI: 01.ijsip.01.01.01
ACEEE International Journal on Signal and Image Processing Vol 1, No. 1, Jan 2010

(                 2
)
here T = α B + 1 − α C ,where B is the same matrix as                                Where, x ( n 1 , n2 ) is the 2D matrix with 0 ≤ n 1 < N1

used in MACE filter and α is some constant which is                                & 0 ≤ n 2 < N 2 and FC ( x ( k 1 , k 2 ) ) represents the 2D-
determined by the constraints of the problem. The value                            DCT        with           0 ≤ k 1 < N1 ,             and     0 ≤ k 2 < N2 .
of α determines the relative importance of noise α = 0                             Normalization coefficient w( k ) is given as
to MACE filter. Using the same argument used to solve                                                              ⎧1 /     2        ( n = 0)
the MACE filter given by equation 16, the solution for                                                    w( n ) = ⎨                                      (20)
OTF can be obtained as                                                                                             ⎩ 1          ( n ≠ 0)

(                        )                          The signs of FC ( x ( k 1 , k 2 ) ) are computed as
−1                       −1            −1
FOTF (h ) = T F ( X ) F ( X ) T F ( X )                     Dv
H
(17)

⎧ −1 ( z (i, j ) < 0)
⎪
sgn ( z (i , j ) ) = ⎨ 0 ( z (i , j ) = 0)
III. DCT SIGN ONLY REPRESENTATION AND DCT SIGN
(21)
ONLY CORRELATION (DCTSOC)
⎪ + 1 ( z (i, j ) > 0)
DCT is a key technology for coded images can be                                                              ⎩
considered as a special case of DFT where in the phase                               For given image x and a template                           g with their
component is zero and the structural information present
in the phase part of the DFT is contained in the sign of                           respective DCT coefficients FC ( x) , FC (g )                  normalized
DCT coefficients. Similar to the concept of Fourier phase                          correlation surface can be expressed as,
only correlation, the idea behind DCT signs only
correlation is to use the important information about the                                                      FC ( x) ⋅ FC (g )
features and details in an image at reduced representation                                           RC =                                                     (22)
FC ( x) ⋅ FC (g )
coefficients (called DCT signs) is robust against scalar                             But           since        DCT          coefficients         are         real
quantization noise because positive signs do not change
numbers x / x = sgn( x ) , so the above equation can be
to negative signs and vice versa. Moreover, the concise
expression of DCT sings saves physical space to calculate                          written as,
and store them. Because of these DCT properties, target
FC ( x)        FC (g )
image search and retrieval taking advantage of the DCT                             RC =               ⋅             = sgn ( FC ( x) ) ⋅ sgn ( FC (g ) )
signs in coded image has been studied [10, 11]. It should                                  FC ( x)        FC (g )                                             (23)
be noted that the intelligibility of the sign-only
representation depends on the magnitude “smoothness”                                                       = FSC ( x ) ⋅ FSC ( g )
of the signal being looked at. Since most natural images
The correlation surface/plane of x, g in spatial domain
contain mostly low frequency content, their magnitude
rolls off quickly at high frequency and this leads to the                          is obtained by computing the inverse 2D-DCT of R                       C
ie.
situation where the “high pass” interpretation of the
IFC ( R C ) .
phase-only transform holds. Hence transformation into a
phase-only or sign only image can also be approximately
interpreted as a high pass filtering operation.                                       IV. FACE RECOGNITION WITH OTF USING DCTSOC
Unlike DFT which transforms real input into complex                                                  (OTDS)
coefficients, DCT transforms real inputs in to real
For better performance and speed of the real-time
coefficients. There are four different types of DCT, type
systems with hardware constrains where correlation
II is the one which is commonly used for image coding.
filters are used for biometrics, one of the biggest
The 2D version of DCT type II transform and its inverse
bottlenecks in the system’s real time performance is the
is given as
computationally intensive evaluation of the 2-D DFT of
(           )
2
FC x(k 1 , k2 ) =        w(k 1)w(k2 )∑∑ x(n1 , n2 ) ⋅                           images not only for the filter design but also for fast
N                   n1 n2                                  computing of correlations during testing phase [12]. Fast
(18)       algorithms based on two bit quantization of DFT
⎛ ( 2n1 + 1) k1π ⎞ ⎛ ( 2n2 + 1) k2π ⎞                                       coefficients to represent phase information of 2D images
cos ⎜                ⎟ ⋅ cos ⎜          ⎟                                       termed as Quad Phase MACE (QP-MACE) filters have
⎝ 2N1 ⎠ ⎝ 2N2 ⎠                                                             been suggested [13].
In general, the MACE filters and OTF make use of
x(n1 , n2 ) = IFC ( x(k 1 , k2 )) =
2
w(n 1 )w(n2 )                            DFT for correlation as explained in section 2. We
N                                            propose to use correlation using DCT sign only values.
(19)
⎛ ( 2k + 1) n π ⎞ ⎛ ( 2k + 1) n π ⎞                          Using the properties of DCT as explained in section 3 the
∑∑ x(k 1 , k2 ) ⋅ cos ⎜ 1 1 ⎟ ⋅ cos ⎜ 2 2 ⎟                                        design equation for standard OTF using DCT sign only
k1 k2                 ⎝ 2N1 ⎠ ⎝ 2N2                     ⎠                          representation (OTDS) can be written as

3
DOI: 01.ijsip.01.01.01
ACEEE International Journal on Signal and Image Processing Vol 1, No. 1, Jan 2010

only representation

(
FSC ( Xk ) = sgn FC ( Xk )      )
2. Design the OTDS filter using DCT sign only values:

(               (
FOTDS (hk ) = sgn T−1FsC ( Xk ) FsC ( Xk )T T−1FsC ( Xk )   )
−1
Dv   )
Figure 1. Correlation outputs of OTDS filter designed using images
of the first person from AMP expression database, when tested for                                        3. Build a filter bank consisting of one filter per
(A) a sample test image of the first subject (B) a sample test image of                                subject:
second subject
(
FOTDS (H ) = FOTDS (h1 ),..., FOTDS (h M )   )
(   −1
(
FOTDS (h ) = sgn T FsC ( X ) FsC ( X ) T FsC ( X )
T     −1
)
−1
)
Dv (24)
Algorithm Testing :
1. Read the testing image to be verified: y

(
T = αB +                1 − α C ,where
2
)                                                 2. Obtain DCT Sign only representation:
FSC (y ) = sgn ( FC (y ) )

(                                                 )
M
1
B = diag           ∑ ( Fs        C
( x i ) FsC ( x i ) )                   (25)
3. Cross correlate the DCT sign only representation
with each of the filters in the designed filter bank and
M   i=0
compute the PSR from the correlation output of each of
Fig1 shows the correlation outputs of OTDS filter                                                    the filters in the filter bank.
designed using all the images of the first person from
AMP expression database, when tested for a sample test                                                                    PSR (k ) = FOTDS ( H ) ⋅ FsC ( y )
image of the first subject and a sample test image of
4. Identify the test image as the subject with max PSR.
second subject.
V. EVALUATION METHOD
One of the most commonly used similarity measure in
correlation filters is based on the value of the largest                                                   In order to have a measure of representation of the
correlation peak as the match score metric. However, to                                                 training dataset by the code book vectors obtained from
minimize sensitive of the match-score metric to variations                                              the algorithm the following five tests are conducted.
such as illumination change and shift. A metric termed as
A. Method I (In-Database test):
peak-to-sidelobe ratio (PSR) which measures the peak
sharpness of the resulting correlation plane is used. The                                                  To test the learning ability, we use all the images
PSR is computed by masking out a 5X5 rectangular                                                        present in a given database as the training set. This gives
region centered at the peak and the defining the                                                        a measure of learning ability of the algorithm for a given
remaining annular region as the sidelobe region. The PSR                                                database in terms of identification rate when all images
metric defined as,                                                                                      present in the training set are used for testing.
B. Method II (Out-Database test):
peak − mean
PSR =                                                                     (26)          To test the ability of the algorithm to identify known
σ                                                             faces with unknown variations i.e. expression unseen
where peak indicates the maximum correlation output,                                                 during training, we use the hold out method/leave one out
the mean and standard deviation (σ ) comes from the                                                     method       by excluding all images with identical
side-lobe region surrounding the peak region.                                                           expression from the training set and use this hold out
images/expression from each of the individual as the test
set. Analysis of the identification rate of this test reveals
A. Proposed Approach                                                                                    performance of the algorithm to identify subjects for a
For a given set of N training images per subject the                                                  particular expression unseen during training
design procedure for the proposed algorithm can be                                                      C. Method III (3 Fold Best):
generalized as given below
In this model of evaluation, approximately one third of
Input :                                                                                                 the total expressions in a given data base with best
Read the training set of N images for each of the M                                                   identification rate during leave one out method (Method
subjects                                                                                                II) are considered for training. And the test set consisted
of all the images belonging to the rest of the expressions.
Xk = {x1,..., xi }        1≤i≤ N
k = 1,..., M                                 Note that the set up of our 3 fold method is different from
the normally used 3 fold method, where two third of
Algorithm Training :                                                                                    entire dataset is used for training and one third of the
1. Obtain the bipolar representation from DCT Sign                                                    dataset is used for testing.

4
DOI: 01.ijsip.01.01.01
ACEEE International Journal on Signal and Image Processing Vol 1, No. 1, Jan 2010

D. Method IV (3 Fold Least):
This model of evaluation is similar to that of method
III but instead of using expressions with best
identification rate for training we use the expressions                                                            TABLE II.
associated with least identification rate during the hold-
AVERAGE RECOGNITION RATE (%) OF STANDARD OTF AND OTF-
out/out-database test method as the training set.                                        DSOC, USING EACH OF THE FIVE EVALUATION METHODS ON AR
FACE DATABASE.
E. Method V (3 Fold Interim):
Evaluation methods
In this 3 fold mode of evaluation, the training set
consists of one third the total number of expressions, with                                          Method      Method      Method      Method   Method
one third of expressions in the training set are drawn from                                            I           II         III         IV        V
the training set of Method III and Method IV and the rest                                Standard
of the expressions where those which exhibited                                                         98.6        82.0       68.8        64.4     79.1
OTF
intermediate identification rate during the leave one out             Figure 2. Out-data identification rate of each of the 64 expressions of
OTDS          95.5
Extended Yale-B database obtained using, 83.9           79.0
standard OTF and OTDS  64.4     85.6
test.
approach.
Based on the five evaluation methods as discussed
above we compare the results of standard OTF approach
with the proposed OTDS approach on four standard

TABLE I.
AVERAGE RECOGNITION RATE (%) OF STANDARD OTF AND OTF-
DSOC, USING EACH OF THE FIVE EVALUATION METHODS ON
EXTENDED YALE-B FACE DATABASE.

Evaluation methods

Method    Method      Method      Method   Method
I         II         III         IV        V
Standard
89.0       86.3        74.9       55.7      84.1
OTF

OTDS        89.4       86.1        78.4       55.7      83.4

Figure 3. Images of the first subject from Extended Yale-B face
frontal face datasets of different size and varied                    database arranged sequentially (top left to bottom right) in terms of
constraints imposed on the acquired face images. For                  the expression index 1-64.
computational efficiency the images in all the datasets
were down sampled to size nearly 32x32 pixels by                     with high identification rate during the hold out method)
maintaining their aspect ratio.                                      an improvement of nearly 4% is observed for OTDS
approach over standard OTF approach.

VI. SIMULATION RESULTS                            B. Yale face database
A. Extended Yale-B face database                                        The Yale database from Yale centre for Computational
Vision and Control contains 165 frontal face images
The extended Yale Face Database B contains 16128                  covering 15 individuals taken under 11 different
images of 37 human subjects under 64 illumination                    conditions; a normal image under ambient lighting, one
conditions. Figure.3 shows expressions/images of one of              with or without glasses, three images taken with different
the subjects.                                                        point light sources, and five different facial expressions.
To verify the performance of the proposed method to               Figure 4 illustrates images from a sample subject from
extreme lighting variations, we have conducted                       this database
experiments on cropped version of extended Yale-B                       A comparison of identification rates between proposed
database scaled to a size of 32x32 pixels. Table 1 gives a           OTDS approach and standard OTF approach for Yale
comparative list of experimental results obtained using              dataset is listed in Table 2 and Figure 5. The
the five evaluation methods discussed in section 3.2.                improvement in performance of the OTDS approach for
Figure 2 gives details of identification rate of each of the         evaluation method II-V indicate that when the constrain
expression using the holdout-expression/out-database                 on representation of frontal face by a mug-shot/cropped is
method.                                                              relaxed and when one allows for variations such as hair
For the evaluation methods as given in section 5.1 the            style and head gear to be present in the training/test set
identification rate obtained for Extended Yale-B dataset             the proposed OTDS approach is able to provide a more
using OTDS approach were almost identical to the                     generalized filter representation.
standard OTF approach, However it should be noted that
in case of evaluation method III when the filters were
designed with better expressions/images (i.e. expressions
5
DOI: 01.ijsip.01.01.01
ACEEE International Journal on Signal and Image Processing Vol 1, No. 1, Jan 2010

Figure 4. Images of the first subject from Yale face database arranged
sequentially (top left to bottom right) in terms of the expression index 1-
11 (Normal, Glasses, Happy, Left Light, Center Light, Surprised, Right

Figure 7. Out-data identification rate of each of the 13 expressions of
AR face database obtained using standard OTF and OTDS approach

The AR face database created by Alex Martinez and
Robert Benavente consist of 126 subjects; 70 male and 56
female with 13 images per person with different
variations in expression, illumination conditions and
occlusions. Figure 6 depicts sample image from one of
the subjects from this dataset.
To further demonstrate the performance of the
proposed approach experiments are conducted using AR
Figure 5. Out-data identification rate of each of the 11 expressions of           face database. The results illustrated in Table 3 and
Yale database obtained using standard OTF and OTDS approach.                      Figure 7 confirm the robustness of the OTDS approach
over the standard OTF approach for datasets with
C. AR face database                                                              variations such as expression and head gear.
D. AMP expression face database
This dataset has 13 subjects each subject being
represented with 75 images showing different
expressions. These face images are collected in the same
lighting condition using CCD camera and have been
TABLE III.
well-registered by their eyes location. Figure 8 shows
AVERAGE RECOGNITION RATE (%) OF STANDARD OTF AND OTF-                             some expression images of one subject
DSOC, USING EACH OF THE FIVE EVALUATION METHODS ON YALE
FACE DATABASE.
Experimental results as tabulated in Table 4 indicate
that both approaches performed extremely well on the
Evaluation methods                              AMP expression dataset and there is nothing much to
Method      Method      Method       Method      Method              choose from as far as this dataset is considered
I           II         III          IV           V
Standard
99.3        80.6        75.1        49.0         80.0
OTF

OTDS          93.3        84.2        81.8        72.1         89.6

Figure 8. Sample images of the first subject from AMP expression face
database

.
Figure 6. Images of the first subject from AR face database arranged
sequentially (top left to bottom right) in terms of the expression index 1
-13 (Neutral expression, Smile, Anger, Scream, Right light on, Left
light on, All sides lights on, Wearing sun glasses, Wearing sun glasses
and right light on, Wearing sun glasses and left light on, Wearing sun
glasses and all light on Wearing scarf, Wearing scarf and left light on,
Wearing scarf and right light on).

6
DOI: 01.ijsip.01.01.01
ACEEE International Journal on Signal and Image Processing Vol 1, No. 1, Jan 2010

REFERENCES
TABLE IV.
[1] A. VanderLugt, “Signal detection by complex spatial
AVERAGE RECOGNITION RATE (%) OF STANDARD OTF AND OTF-                              filtering ,” IEEE Trans. Informatiom Theory ,
DSOC, USING EACH OF THE FIVE EVALUATION METHODS ON AMP
EXPRESSION FACE DATABASE.                                                          no.10,pp.139-145,1964 .
[2] B.V.K. Vijay Kumar, “Tutorial survey of composite filter
Evaluation methods                                 design for optical correlation,” Applied Optics, vol. 31, pp.
4773-4801, 1992.
Method     Method      Method      Method     Method
[3] B.V.K. Vijay Kumar, M. Savvides, C. Xie, K.
I          II         III         IV          V
Venkataramani, J. Thornton, and A. Mahalanobis,
Standard                                                                           “Biometric Verification with Correlation Filters,” Applied
99.6        99.6       95.7        97.1        97.1
OTF                                                                                Optics, Vol 43, No.2 , pp. 391-402, 2004.
[4] P. Refregier, “Filter design for optical pattern recognition:
OTDS         99.3        99.3       96.6        97.1        97.0
multi-criteria optimization approach,” Opt. Letters, vol.15,
pp.854-856, 1990.
[5] A. Mahalanobis, B.V.K. Vijay Kumar, and D. Casaaent,
“Minimum average correlation energy filters,” Applied
Optics, vol. 26, No. 17, pp. 3633-3640,1987.
[6] C. F. Hester, and D. Casasent, “Multivariate techniques for
multiclass pattern recognition,” Applied Optics, vol. 19, pp.
1758-1761, 1980.
[7] B.V.K. Vijay Kumar, “Minimum variance synthetic
discriminant functions,” J. Opt. Soc. Am, vol. 10, No.13,
pp. 1579-1584,1986.
[8] M. Savvides, and B.V.K. Vijay Kumar, “Efficient Design
of advanced correlation filters for robost distortion tolerant
face recognition,” Proceedings of IEEE Conference on
Advanced Video and Signal based Surveillance, 2003,
pp.45-52.
[9] K. Takita, T. Aoki, Y. Sasaki, Higuchi, and K. Kobayashi,
“High Accuracy Sub-pixel Image Registration Based on
Figure 8. Out-data identification rate of each of the 75 expressions of            Phase Only Correlation,” IEICE Trans. Fundamentals,
AMP Expression face database obtained using standard OTF and OTF-                  vol.E86-A, No.8,2003.
DSOC approach                                                                 [10] C. D. Kuglin, and C. D. Hines, “The Phase Correlation
Image Alignment Method,” Proc. International
Conference on Cybernetics and Society, 1975,pp. 163-165.
CONCLUSIONS                                         [11] W. S. Hoge, A Subspace Identification Extension to the
Phase Correlation Method. Harlow, England: Addison-
In this paper, we have proposed and evaluated a DCT                             Wesley.
based version of OTF filter using DCT sign only                               [12] K. N. Chee, M. Savvides, and K. K. Pradeep, “Real-Time
correlation which is efficient both in terms of                                    Face Iderification System on a Cell-Phone using Advanced
computation and storage as it deals with bipolar                                   Correlation Filters,” AutoID, 2005, pp. 69-74.
representation of images and filter coefficients in DCT                       [13] Savvides, M. and Vijay Kumar, B.V.K. “Quad Phase
domain making computation of correlation easier/faster.                            Minimum Average Correlation Filters for Reduced
Based on comparison of the identification rate of the                              Memory Illumination Tolerant Face Authentication,”
Audio-and Video-Based Biometric Person Authentication,
proposed OTDS approach with the standard OTF
4th International Conference, 2003, pp. 45-52.
approach using four standard frontal face database
conclusions can be drawn that:
The performance of OTDS approach is comparable to
standard OTF approach for datasets consisting of mug-
shot frontal faces with only illumination and expression
variations. OTDS approach is as tolerant to illumination
variations as the standard OTF approach, demands less
storage requirement and is computationally less complex
during testing as it makes use of correlation between
bipolar vectors as similarity measure for classification.
For datasets such as AR and Yale, where many of the
images were present with occlusions (scarf and black
sunglasses) OTDS achieved better results than the
standard OTF approach and is hence a better alternate to
standard OTF approach for frontal face recognition .

7