FACE RECOGNITION USING WAVELET REPRESENTATIONS OBTAINED FROM

DIFFERENT PRUNING STRATEGIES

Kenny Teng, James Auwaerter
Department of ECE
Carnegie Mellon University
Pittsburgh, PA 15213
combined tree that is able to perform well at both
verification and classification.

ABSTRACT
We present a comparison between two major pruning
strategies for face recognition using wavelet packets. The
first approach is top-down, beginning with the image
representation of the human faces and decomposing the
image selectively based on correlation energy. The second
approach begins with the full wavelet decomposition,
designs a correlation filter for each subspace, and prunes
the tree based on classification rate. Each technique has
advantages: top-down is better at verifying whether an
image belongs to a particular class while the full wavelet
decomposition is better at determining which class an
image belongs to out of a set of different classes. We
attempt to generate trees that take advantage of both
strengths to see at what middle ground we maximize
effectiveness at image verification and image
classification, using faces as the test images. Then, we
apply the tree structure to sets of human faces to
determine whether the tree structure is more broadly
applicable.
1.

1.2. Previous Work and State-of-the-Art

Previous work in this area includes: Hennings et al. [1]
who experimented with correlation filters as the pruning
strategy for fingerprint classification and Phillips [2] who
designed two sets of filters for his work with face
verification. Each filter set was designed with a single
goal in mind: verification or classification. These filters
were then tested on three different types of face image
sets and compared across the types.
1.3. Short Overview of the Paper
We begin by introducing the basics of biometrics, wavelet
transformations, and classification methods in Section 2.
Our figures of merit and proposed methodology are
described in Section 3. We then describe our results in
Section 4. We end this paper with a brief conclusion.
2.

INTRODUCTION

BACKGROUND

2.1. Biometrics
Biometrics is the use of unique physical characteristics of
people to positively authenticate a user of a system or
recognize a type of a pattern and then identify people
based on that pattern. Examples of biometrics for
authentication include iris pattern, hand vein pattern, and
fingerprints, while biometrics for recognition are
primarily based on facial features. By identifying which
characteristics are particular to a person, that person can
be recognized in the future.
Currently, biometrics is not commonly used to
authenticate users. The most common use, fingerprint
scanners on laptops, is still limited to after-market
peripherals and a few models of IBM laptops [3]. Two of
the problems that prohibit their more widespread use are
the difficulty of obtaining a clean image and the greater
space that it takes to store an image as opposed to an
alphanumeric password. While we can assume that a user
will try to give a clean image in order to maximize
recognition, different lighting conditions or fingerprint
pressure can greatly affect the image that is compared to

1.1. Motivation
While it is a relatively simple matter to either identify a
human face or classify between different ones, it would be
very interesting to see if we could use wavelet transforms
to achieve better and more robust results. One of most
characteristics of wavelet transforms is their ability to
represent a signal into partitions of time-frequency plane.
The popular representation of wavelet transforms is a
multi-resolution wavelet tree where the each sub space
contains information in a time-frequency domain.
Therefore, we want to design figures of merit to take
advantage of those wavelet spaces to end up with different
pruned trees. The main motivation is to try to design a tree
that is optimal for verification and one which is good for
classification of face images. Their performance at
verification and classification will then be measured using
the figures of merit, and the strategies will be refined
based on the results. In the end, we will come up with a

the database image. Authentication is composed of two

parts: recognition that an image is being scanned, and
classification of that image. For the most part, we can
assume that the challenge lies in the classification of the
image.

the original input and filtering it with a combination of

high-pass and low-pass filters, designed to maximize the
amount of information obtained within each subspace.
This decomposition can be repeated for n-levels. The
image can later be reconstructed from these subspaces. By
removing subspaces that contain comparatively low
amounts of information from a reconstruction, we can
achieve an image that is nearly as good as the original but
takes less space to store. This can be useful if we are
storing a large number of similar images.

2.2. Wavelet Transformations

Wavelet transformations are a method of representing
signals across space and frequency. The signal is divided
across several layers of division in space and frequency
and then analyzed. The goal is to determine which
space/frequency bands contain the most information about
an images unique features, both the parts that define an
image as a particular type (fingerprint, face, etc.) and
those parts which aid in classification between different
images of the same type.
One type of discrete wavelet transform (DWT) is the
orthogonal DWT. The orthogonal DWT projects an image
onto a set of orthogonal column vectors to break the
image down into coarse and fine features.

2.3. Classifier
There are many different classifiers out there that have
proved to be very effective in classifying faces. We will
be using advanced correlation filters, specifically the
Minimum Average Correlation Energy filter (MACE).
Correlation filter techniques are attractive candidates for
the matching needed in face verification. According to
Kumar et al. in [4], correlation filters can be used on any
biometrics as long as they are in the form of images.
Advanced correlation filters can offer a very good
matching performance in the presence of variability such
as facial expression and illumination changes.
Furthermore, they are of less complexity and are shift
invariant.
The main idea is to synthesize a filter using a set of
training images that would produce correlation output that
reduces the correlation values at locations other than the
origin and this value at the origin is constrained to a
specific peak value. When the filter is correlated with a
test image that is authentic, the filter will exhibit sharp
correlation peaks in the correlation plane. Otherwise the
filter will output small correlation values. This correlation
plane v can be modeled as

Figure 1 A two-channel filter bank from [5].

In Figure 1, we see the order in which filters are applied

to achieve a simple one-level wavelet decomposition. The
filter h is a high-pass filter and the filter g is a low-pass
filter.

v i = (DFT)1 Ci DFTx

(1)

where x is the vectorized form of the n image pixels, DFT

is an n x n matrix containing the basis of a twodimensional discrete Fourier transform and Ci is a
diagonal matrix containing the correlation filter values for
class i along the diagonal.
From the correlation plane, we can use a match metric
from [1] by measuring the peak-to-correlation energy
(PCE), which is as follows:

PCE(v i ) =
Figure 2 - An example of a full wavelet decomposition with two
levels.

max(| v i |) mean(| v i |)
stdev(| v i |)

(2)

The MACE filter seeks to minimize the average

correlation plane energy:

A typical two-level full wavelet decomposition is shown

in Figure 2. Each of the subspaces is obtained by taking

E average =

where

1
N

Ei =
i =1

1
N

h
i =1

Therefore, the higher the fitness metric, the better the

correlation filter is expected to perform. For a fourchannel wavelet filter bank where a space V0 is
decomposed into four spaces, LL, LH, HL and HH, we
would want to evaluate which level is performing better.
In this case, we form the following inequality

Di h = h + Dh , (3)

1 N
D = Di
N i=1

(4)

F(V0 ) > F(LL) + F(LH) + F(HL) + F (HH)

Minimization of Eaverage is done while satisfying the linear

constraints of the correlation values based on prespecified values , that is in row vector u,

X +h = u

where the operator F(.) computes the fitness metric of the

space. Here, LL, LH, HL and HH are the spaces of signals
x (ggi +1) , x (ghi +1) , x (hgi +1) and x (hhi +1) respectively. If the left-

(5)

hand side of the inequality is greater, we consider it as a

better space to generate a correlation filter, and if the
right-hand side is bigger, then four separate filters will be
generated instead.

where X is a d N complex matrix, with the ith column

containing the 2-D Fourier transform of the ith training
image lexicographically re-ordered into a column vector.
Lagrange multipliers method is used to minimize (3)
while satisfying the linear constraints in (5) and this will
give a closed form solution for the MACE filter as
follows:
2

h = D 1 X ( X + D 1 X ) 1 u
3.

(9)

3.1.2. False Positive Rate

The second figure of merit is based on the fact that we are
trying to optimize the tree for classification. Intuitively, by
measuring the false positive rate that a given correlation
filter yields, we can define a measure of performance. A
false positive is recorded whenever, a test image is said to
belong to a class other than its real class. Therefore, for a
given space, V0 and its corresponding correlation filter,
the PCE value of all the classes other than the class it
belongs to, are computed. Then, the performance
inequality is as follows:

(6)

PROPOSED WORK

3.1. Figures of Merit

We used two different figures of merit to evaluate the
performance of the wavelet trees. The first one is based on
the correlation energy of the filters at different wavelet
sub-spaces and the other one is the false positive rate
achieved by a wavelet-domain correlation filter.

max(PCE (Vi0 )) > E(PCE (V0 ))

(10)

where 0 < < 1 and E(.) operator is the expected value of

the PCE value for space V0. If the above inequality is
satisfied, this means that the false positive rate of this
correlation filter tends to be high, and therefore the
discriminative power of that filter is not good enough.

3.1.1. Correlation Energy

According to Savvides in [7], if we use correlation filters
as our main classifier, then a good figure of merit would
be to measure the peak-to-sidelobe ratio of the correlation
output to each subspace of each image. By substituting (6)
into (3), we end up with:

3.2. Match Metric

In the evaluation stage, we need to compute a match
metric when the wavelet-domain correlation filters are
applied to a fully decomposed image. Therefore, after
decomposing an image into a wavelet tree, the waveletdomain correlation filters are applied to generate the
correlation planes. Here, we can extend the PCE
computation from (2) to calculate the sum of all the PCE
values of all the wavelet sub spaces.
The formulation of this calculation is as follows. For a
given image x, and wavelet packet correlation filters of
class i contained in matrices Ci,l for every subspace l, the
subspace correlation places are given by

E = (D1 X(X + D1 X )1 u) + D(D1 X (X + D1 X)1 u)

= uT (X + D1 X)1 u
(7)
where X holds the spectrum of one of the training images
after projecting the image onto the wavelet subspace in
each column. Since better performance is obtained by
lower energy, then we define our fitness metric as
~

F = 1/ E

(8)

v i,l = (DFT)1 Ci,l DFT WPi,l x

(11)

and the match metric becomes

f (v i ) = PCE(v i )
l

(12)

3.3. Methodology
We will be using the Cohn-Kanade AU-Coded Facial
Expression database for testing our pruning strategies. For
the purpose of verification and classification, Advanced
Correlation filters, more specifically MACE will be used.
The evaluation process will consist of two phases, namely
training followed by testing phase. For the first stage, the
first tree structure is going to be designed by starting in
the image space and adaptively decomposing using
wavelet transforms. At the same time, pruning will be
done at every sub space using correlation energy based on
inequality (9) as main figure of merit. The second strategy
is that we start with the full decomposition and prune
every single sub space using the second figure of merit
defined in (10). The design of those two tree structures
will be adaptive whereby we can alter the variables in our
figure of merits to get those two representations as close
as possible. In other words, we designed an optimal tree
with the best verification rate and at the same time
requiring the minimum spanning tree. The way to do that
is that we build a combined tree, which has a correlation
filter at a sub space on both trees. Once we have the
combined tree, we can begin the testing phase. We tested
it by applying those wavelet-domain correlation filters on
the expression database to evaluate the effectiveness of
our method. As a control test, standard correlation filters
have been generated in the image-intensity domain only
and their overall performances for verification and
classification have been computed. We will compare our
results to those values.

4.

Figure 3 An instance of a particular class from the

expression database.

In the database we used, there are 20 classes (persons)

with 20 different faces each, with variations facial
expressions.
4.2. Evaluation
For verification evaluation of our method, a receiver
operating characteristics (ROC) curve has been plotted to
show the probabilities of authentic versus the probabilities
of impostors. From the following graph, it is clear that our
optimized tree achieved better results (blue line) than the
control test (red line). The probability of having an
authentic verification with zero probability of having an
impostor is about 83% for wavelet-domain method and
67% for the standard filters.

EXPERIMENTAL RESULTS

4.1. Datasets
The database we used for evaluation is the Cohn-Kanade
AU-Coded Facial Expression database.

Figure 4 ROC curve showing the results for verification

for both wavelet-domain and standard correlation filters.

As for classification results, the error rates have been

plotted for every single class as shown in figure 5. There

[1]

is a clear reduction in classification error rate for the

wavelet-domain filters (red) compared to the standard
correlation filters. For class 4 where the error rate for the
control test reaches 70%, it is still hard for even the
wavelet-domain filters to classify efficiently, even though
they yielded a better rate of 65%. However, for classes
like 13 and 17 where the standard method had error in the
order of 35% and 10% respectively, the wavelet-domain
method achieves a zero error rate. For the overall
classification rate, our wavelet-domain method yields
94.2% accuracy while the standard method achieves 90%.
Classification Error Rate

[2]
[3]
[4]
[5]

Original
Wavelet

80

[6]

70
60

[7]

50
40
30
20

[8]

10
0
1

9 10 11 12 13 14 15 16 17 18 19 20
Classes

Figure 5 Classification error rate of both wavelet-domain

and standard correlation filters

CONCLUSIONS
We have discussed the basic elements of biometrics and
wavelet transforms, and how correlation filters may be
used to classify images within a biometric system. We
explored the advantage of using wavelet packet
decomposition for verification and classification and
determined how to best use our figures of merit to obtain
an optimal wavelet decomposition tree. The combined
wavelet tree performed better than the standard correlation
filters applied only in the image-intensity domain. The
results show that face images have some features that
remain more consistent in the wavelet sub spaces than in
the spatial domain. Future work may involve designing
different figure of merits that will be tailored to specific
datasets.
ACKNOWLDGEMENTS
The author would like to thank Jelena Kovaevi, Pablo
Hennings, and Amina Chebira for their help in conducting
the experiments and the class of 18-795 for their feedback
during the class and presentations.
REFERENCES

P. Hennings, J. Thornton, J. Kovaevi, and B. Kumar,

Wavelet packet correlation methods in biometrics
Journal of Applied Optics, special issue on Biometric
Recognition Systems.
Phillips, P.J., Matching pursuit filters applied to face
identification. IEEE Transactions on Image Processing,
Volume 7, Issue 8, Aug. 1998 Page(s): 1150 1164
IBM Introducing Fingerprint Reader into Laptop;
Germain, Jack: Written 10/04/04, Accessed 12/01/05;
http://www.technewsworld.com/story/37017.html
B. V. K. Vijaya Kumar, A. Mahalanobis, S.R.F. Sims, J.F.
Epperson, "Unconstrained correlation filters," Applied
Optics, vol. 33, pp. 3751-3759, 1994.
A. Chebira, G. Srinivasa, J. Kovaevi, A wavelet
packets approach to classification of subcellular protein
images. Unpublished work, October 2004
M. Vetterli and J. Kovaevic, Wavelets and Subband
Coding, Prentice-Hall Signal Processing Series, 1995.

Marios Savvides, B.V.K. Vijaya Kumar, P.K.

Khosla, "Illumination Tolerant Face Identification using
Advanced Correlation Filters trained from a single face
image," accepted for publication in Biometrics
Consortium Conference BC2004.
R. Duda, P. Hart, D. Stork, Pattern Classification, Wiley,
New York, 2000