Face Recognition based on Discrete Wavelet
Transform and Sobel - Roberts Features
Mahdi parhizkari hossein khosravi
School of Electrical Engineering School of Electrical Engineering
Shahrood University of Technology Shahrood University of Technology
Iran, Shahrood Iran, Shahrood
Parhizkari.mahdi@gmail.com hosseinkhosravi@gmail.com
Abstract— Face is one of the physiological features used in
identifying or verifying a person. Different methods have been
proposed for recognition of a person’s face in a pool of images
of faces belonging to different persons. We used Discrete
Wavelet Transform (DWT) and Sobel-Roberts Feature (SRF)
for face recognition. For this purpose, in the training phase,
DWT was applied to the input image, and SRF was used to
extract features from approximation image. The system was
trained through Support Vector Machine (SVM) and the
parameters were stored. In the recognition phase, again
features of the input image were extracted and faces were
recognized using SVM and the parameters stored in the
previous stage. Olivetti Research Laboratory (ORL) database
was used for the purpose of recognition in this study. For 4 and
5 images of any person for the training set gave recognition
rates of 97.5% and 98.6%, respectively which is higher than
other methods. The time required for phase detection was
approximately 7ms. Results show that the proposed method
has high accuracy and high speed.
Keywords- Face Recognition, Discrete Wavelet Transform
(DWT), Sobel-Roberts Features (SRF), Support Vector Machine
(SVM)
I. INTRODUCTION
During the recent years, face recognition has drawn
attentions in areas related to biometrics and pattern
recognition [1] [2]. Different methods have been proposed
for recognition of two-dimensional facial images. These
methods can be classified into three categories:
a) Feature-based approaches to face recognition among Turk
et al. (1999) proposed eigneface recognition which applies
principle component analysis (PCA) to reduce dimensions.
They found a subspace with orthogonal vectors used to
express images in this subspace and to obtain weights that
represented images in this space. They then compared stored
weights to the weights obtained for the new image to
recognize this image [3]. b) In techniques based on models,
features, or texture of certain parts of face among Wiskott et
al. (1997) presented elastic bunch graph which bases
recognition on similarities found between two image graphs.
First, key spots, such as nose, eyes, eyebrows, etc , were
identified and then Gabor wavelets we applied to extract
features for each key point. The graph nodes included these
key points, each node representing Gabor jets extracted from
these points. Similarities between two images were found
based on graph similarities [4]. Cootes et al. (1998) proposed
978-1-4673-6206-1/13/$31.00 ©2013 IEEE
hadi grailu
School of Electrical Engineering
Shahrood University of Technology
Iran, Shahrood
grailu@gmail.com
active appearance model which combined appearance and
facial features [5]. c) Some techniques combined eigenface,
eigennose, and eigeneye [6].
Zhao et al. conducted an exhaustive review on 2D face
recognition [7].
The rest of this paper is organized as follows: in Section II,
describes the proposed facial recognition system. In Section
A, Discrete wavelet transform (DWT) is introduced and how
to obtain approximate image using input image is expressed.
In Section B, Sobel-Roberts Features (SRF) is introduced
and feature extraction of image approximation in a different
direction is expressed. In Section III, using Support Vector
Machine has been identified and the results obtained using
the ORL database has been placed.
II. PROPOSED SYSTEM
As seen in Figure (1), the proposed recognition system
divides the data into two classes: test data and training data.
In training phase, first DWT was applied to the input image,
and SRF was used to extract features from approximation
image. The system was trained through Support Vector
Machine (SVM) and the parameters were stored. In the
recognition phase, again features of the input image were
extracted and faces were recognized using SVM and the
parameters stored in the previous stage.
A. Discrete Wavelet Transform (DWT)
For a single-dimension signal, a low-pass and a high-pass
filter can be used together with decremental sampling with a
ratio of 2 to create high-frequency and low-frequency
regions. Number of sample in each region is half the number
of samples in the original signal. The low-frequency region
is further divided into two sub-regions with high and low
frequencies and decremental sampling can again be used to
reduce the number of samples to one fourth the samples in
the original signal. The decomposition level for wavelet
transform is represented by the number the abovementioned
procedure is repeated.
For 2D signals, the same procedure is applied once in
horizontal direction and another time in vertical direction [8].
In the first step, low-pass and high-pass filters are applied in
succession to the rows of the input image followed by a
decremental sampling with a ratio of 2 leading to an image
with high and low frequencies. In the second step, high-pass
and low-pass filters are applied in succession to
the columns of the two images produced in the previous
 Figure 1. proposed system for face recognition

step. Again, a decremental sampling withh a ratio of 2 is a) A 92×112 image was divided into sixteen 23×28 non-
applied. As a result four images aree produced: the overlapping blocks.
approximation component I LL , the horizoontal component b) For each pixel in each block Sobbel’s operator was used to
I LH , the vertical component I HL , annd the diagonal find the gradient as seen in Table (I). Equations (1) through
component I HH . An example of 2D D Discrete Wavelet (3) were used in the procedure.
Transform on facial image can be seen in Figgure (2).
T
g x, y g ,g where 1

g x, y f x 1, y 1 2f x 1, y

f x 1, y 1 f x 1,
1y 1

2f x 1, y f x 1, y 1 2

g x, y f x 1, y 1 2f x, y 1

f x 1, y 1 f x 1,
1y 1

2f x, y 1 f x 1, y 1 3

Here, f x, y represents the value forr a pixel located at x, y .

c) Equations (4) and (5) were used u to find phase and
magnitude of gradient vector.

Figure 2. 2D Discrete Wavelet Transform For two-sstage on the image of p x, y tan g ,g 4

the ORL database

B. Sobel-Roberts Features (SRF) A x, y g g 5

SRF was first employed by khosravi et al. in recognition of
fonts used in Farsi font recognition. E Empirical results
indicated enhanced accuracy and speed com mpared to Gabor
filter [9]. In this paper we apply SRF to featuure extraction for
ORL images. SRF was designed as follows:
d) Phase angles were quantized into o 16 intervals (from 0 to
2 15 16).Then, Equation (6) was ussed to identify 16 features
corresponding to these 16 phases forr each block.
 III. RESULTS
F A x, y 6 Olivetti Research Laboratory (ORL) database was used in
,
this paper. Figure (3) shows samples of images available in
ORL database [10].
F represent the feature corresponding to θ . In addition,
(x , y ) represents all coordinates with phases equal to θ.
e) In each pixel of each block, gradient was calculated using
Roberts’s operator (Table (II)) and Equations (7) and (8).

g x, y f x, y f x 1, y 1 7

g x, y f x 1, y f x, y 1 8

f) Equations (4) and (5) were used to find phase and

magnitude of the gradient vector.
g) Through iteration of Step 4, 16 features were extracted for
each block.
In sum, 16×16 = 256 features were extracted using Sobel’s
operation and 16×16=256 features were extracted through
Roberts’s operation. The features were then normalized. In
total, 512 features were extracted for 16 rotations. Also 128
and 256 features were extracted for 4 and 8 rotations.

TABLE I. (A) , (B) SOBEL’S OPERATOR Figure 3. An example of images of ORL database
1 0 -1 1 2 1
2 0 -2 0 0 0 To evaluate the results of the proposed method on ORL,
1 0 -1 -1 -2 -1 we utilized four wavelet transforms db1, db2, db3, and db4.
Once DWT was applied to the images, SRF was used for
A B feature extraction in 4, 8, and 16 directions. Tables III and IV
present recognition rates for different numbers of images per
TABLE II. (A) , (B) ROBERTS’S OPERATOR individual in the training set for one-step and two-step
transforms, respectively.
0 1 1 0 The figures presented in these two tables reflect mean rates
-1 0 0 -1
obtained through 5 iterations. Table V shows highest
A B recognition rates with lowest number of features based on
Tables III and IV.

TABLE III. RECOGNITION RATES FOR NUMBER OF DIFFERENT IMAGES OF EACH PERSON FOR ONE-STEP TRANSFORMS

Number of image
Type of wavelet 3 4 5 6 7 8 9
Number of direction
4 93.71% 96.41% 98.30% 99.37% 99.33% 99.5% 100%
Db1 8 92.35% 97% 97.3% 99.37% 99.83% 100% 100%
16 93.14% 95.91% 98.60% 98.5% 99.66% 100% 100%
4 93.78% 96.08% 98.6% 98.37% 99.16% 99.25% 99%
Db2 8 94.07% 96.25% 98.2% 98.87% 99% 100% %100
16 94.85% 96.41% 98.1% 99.12% 100% 99.75% 99.5%
4 92.78% 97% 97.6% 98.87% 99.16% 99.75% 100%
Db3 8 94.35% 95.75% 98.4% 99% 99.5% 100% 99.5%
16 93.71% 96.75% 98.6% 99.25% 99.83% %99.75 99.5%
4 93.44% 96.58% 96.9% 97.75% 99.33% 98.75% 99.5%
Db4 8 94.85% 96.50% 98.40% 99.37% 99.66% 100% 100%
16 94.71% %97.5 98.6% 99.25% 100% 100% 100%
 TABLE IV. RECOGNITION RATES FOR NUMBER OF DIFFERENT IMAGES OF EACH PERSON FOR TWO-STEP TRANSFORMS

Number of image
Type of wavelet 3 4 5 6 7 8 9
Number of direction
4 92.58% 97.5% 97.6% 98.62% 99.83% 98.5% 99.5%
Db1 8 93.57% 96.41% 97.1% 98.12% 99.5% 98.75% 99.5%
16 90.14% 96.08% 97.6% 97.87% 99.16% 99.5% 100%
4 94.5% 95.91% 97.9% 98.75% 98.16% 99.5% 99.5%
Db2 8 92.57% 95.66% 98.2% 98.12% 98.5% 99% 98.5%
16 92.5% 96.08% 97.3% 99.12% 98.5% 99% 100%
4 94.64% 96.33% 98.5% 98.62% 99.83% 99.25% 100%
Db3 8 93.64% 97.08% 98.3% 99.12% 98.66% 99.75% 99.5%
16 93.57% 96.83% 98.2% 98.5% 99.33% 99.25% 99%
4 91.87% 97.33% 98% 98.12% 98.83% 99.25% 100%
Db4 8 94.07% 97.5% 98.6% 99.12% 99.5% 99.25% 99.5%
16 93.42% 96.08% 97.3% 98.87% 99.83% 99.25% 100%

TABLE V. HIGHEST RECOGNITION RATES FOR NUMBER OF DIFFERENT IMAGES OF EACH PERSON ACCORDING TO TABLE IV

Number of image 3 4 5 6 7 8 9
highest recognition rate 94.87% 97.5% 98.6% 99.37% 100% 100% 100%
Minimum number of feature 256 128 128 128 512 256 128

Table (VI) presents recognition rates in different methods

applied to 4 images of each individual within the training set. TABLE VII. RECOGNITION RATES IN DIFFERENT METHODS APPLIED TO
The recognition rates cited for different methods are based 5 IMAGES OF EACH INDIVIDUAL WITHIN THE TRAINING SET
on [11]. As seen in this table, the proposed method has a METHOD RECOGNITION RATE
higher recognition rate compared to other methods when
Contourlet+LDA 78.5%
applied to 4 images of each individual in the training set.
Table (VII) presents recognition rates in different methods ICA 80.5%
applied to 5 images of each individual within the training set.
LDA 87.5%
The recognition rates cited for different methods are based
on [12]. As seen in the table, the proposed method has a PCA 88.5%
higher recognition rate compared to other methods when DWT+LDA 89.5%
applied to 5 images of each individual in the training set.
Standard Eigenface 92.2%
TABLE VI. RECOGNITION RATES IN DIFFERENT METHODS APPLIED TO
Waveletface 92.5%
4 IMAGES OF EACH INDIVIDUAL WITHIN THE TRAINING SET
Waveletface+PCA 94.5%
METHOD RECOGNITION RATE
NCC 84.86% Waveletface+LDA 94.7%

Hidden Markov Model (HMM) 87% Waveletface+LDA+NFL 95.2%

Eigenface 90% Waveletface+KAM 96.6%

SVM 91.21% Curveletface+PCA 96.6%

DCT+GA 94.4% Curveletface+PCA+LDA 97.7%

DCT+PSO FS 94.7% Curvelet +LDA 98%
2D+HMM 95% DWT+SRF+SVM 98.6%
Hybrid NN:SOM+CNN 96.2%
DWT+PSO FS 96.8% For computer with characteristics: Core(TM) i3 CPU
M370 @2.40GHz 2.39 GHz, Table (VIII) shows the time
DWT+GA 97%
required for recognition phase. As seen in the table, as
DWT+SRF+SVM 97.5% number of rotation doubles, recognition time increases by
approximately 1ms. The high speed of the proposed method
is one of its advantages over other methods.

TABLE VIII. TIME REQUIRED FOR RECOGNITION PHASE

number of wavelet
1 level 2 level
number of rotation
4 7.01 s 7.08 s
8 7.10 s 7.21 s
16 7.22 s 7.29 s

CONCLUSION
Face recognition was applied through DWT and SRF.
Four wavelet (db1, db2, db3, and db4) were used in DWT
while three directions (4, 8, and 16) were utilized in the SRF.
The network used in the proposed recognition system was
SVM. The proposed method was implemented on ORL.
Application of the proposed method to 4 and 5 images for
each individual in the training set gave recognition rates of
97.5% and 98.6%, respectively. The recognition phase
duration obtained by this method was about 7ms. The results
indicate that the proposed method is fast and accurate and
can effectively comparable to other methods.
SUGGESTIONS
Many features are undesirable and can be eliminated.
Therefore, we suggest application of PSO-based feature
selection or GA-based feature selection to reduce the number
of features.
REFERENCES
[1] H. Wechsler, et al., "Face Recognition: From Theory to
Applications," presented at the Springer-Verlag, 1966.
[2] S. Gong, et al., "Dynamic Vision : from Images to Face
Recognition," presented at the Imperial College Press and
World Scientific Publishing, 2000.
[3] M. Turk and A. Pentland, "Eigenfaces for recognition,"
Journal of Cognitive Neuroscience, vol. 3, pp. 71-86, March
1991.
[4] L. Wiskott, et al., "Face Recognition by Elastic Bunch Graph
Matching," IEEE Trans. Pattern Analysis and Machine
Intelligence, vol. 19, pp. 775-779, 1997.
[5] T. F. Cootes, et al., "Active appearance models," presented at
the EuropeanConference on Computer Vision, 1998.
[6] Pentland, et al., "View-Based and Modular Eigenspaces for
Face Recognition," presented at the IEEE Computer Vision
and Pattern Recognition, 1994.
[7] W. Zhao, et al., "Face Recognition: A Literature Survey,"
presented at the ACM Computing Surveys, 2003.
[8] S. Burrus and R. Gopinath," Introduction to wavelets &
wavelet transforms," New Jersey, 1998.
[9] H. Khosravi and E. Kabir, "Farsi font recognition based on
Sobel–Roberts features," Pattern Recognition Letters, vol. 31,
pp. 75-82, 2010.
[10] "The ORL Database of Faces," ed. Cambridge University
Computer Laboratory, between April 1992 and April 1994.
[11] R. M. Ramadan and R. F. Abdel-Kader, "Face Recognition
Using Particle Swarm Optimization-Based Selected
Features," International Journal of Signal Processing, Image
Processing and Pattern Recognition, vol. 2, pp. 51-66, June
2009.
[12] M. Rziza, et al., "Local Curvelet Based Classification Using
Linear Discriminant Analysis for Face Recognition,"
International Journal of Electrical and Computer Engineering,
vol. 4, pp. 72-77 ,2009.