http://www.paper.edu.

cn

FEATURES EXTRACTION USING A GABOR FILTER FAMILY

Danian Zheng, Yannan Zhao, Jiaxin Wang
State Key Laboratory of Intelligent Technology and Systems
Department of Computer Science and Technology
Tsinghua University, Beijing 100084, China
email: [email protected]

ABSTRACT eters [4]. In image analysis and compression, Daugman

Gabor filters possess the optimal localization properties in presented a three-layered neural network for transforming
both spatial and frequency domain, and they have been suc- two-dimensional discrete signals into ge- neralized non-
cessfully used in many applications. But how to design a orthogonal 2-D Gabor representations for image analysis,
set of befitting Gabor filters for a specific application has segmentation and compression. The Gabor filters used in
maybe puzzled many users for a long time. In this paper, his paper are based on a biologically inspired log-polar en-
we purpose designing a common set of Gabor filters – a semble of dilations, rotations and translations of a single
Gabor filter family to solve the problem. The Gabor fil- underlying 2-D Gabor wavelet template. The neural net-
ters in the family are well combined to capture the whole work is used for finding the optimal coefficients of Gabor
frequency spectrum in all directions. And we can extract transform [1]. In edge detection, some authors investigate a
many meaningful features using the Gabor filter family. 2-D Gabor odd filter-based detector, and the overall perfor-
Experimental results in textures and characters demonstrate mance of this detector is almost identical to that of the first
these features commendably expressing the local informa- derivative of Gaussian [2]. But differing from them, we
tion with the different frequencies and orientations in the purpose designing a common multi-resolution Gabor filter
image. The Gabor filter family designed by us can be also family used to extract features from the images. In this pa-
used in some other applications. per, two issues are mainly considered: one is to design a
preset common set of Gabor filters for these applications,
KEY WORDS and we call them a Gabor filter family; the other is to ex-
Features extraction, Gabor filters, Gabor filter family, tex- tract features using the Gabor filter family from an image.
ture feature, and character feature After the 1-D Gabor function had been proposed in 1946
by Gabor in his paper ”Theory of communication”, Daug-
man extended it to a 2-D Gabor filter and showed it provid-
1 Introduction ing simultaneous optimal resolution in both the space and
frequency domain. The Gabor filters defined in our paper
In recent years, Gabor filters have been successfully have the equivalent energy 1, and they are all derived from
used in many applications, such as texture segmen- a common basic Gabor filter by rotating and dilating. These
tation/classification [3,4,11], target detection, character Gabor filters are well combined to cover over the whole 2-
recognition [7,8], fingerprint recognition [5,9], face recog- D frequency domain. The features extracted using Gabor
nition [6,10], fractal dimension management, document filters represent the local information in the image. We can
analysis, edge detection [2], image analysis and compres- take features from the amplitude or phase after convolving
sion [1]. Gabor filters have received considerable attention, the complex Gabor filters with the image. In this paper,
because they possess the optimal localization properties in we give out four feature examples. In order to reduce the
both spatial and frequency domain. computational complexity, we can choose the numbers of
Designing some befitting Gabor filters for a particu- Gabor filters (frequency number and orientation number),
lar processing task and reducing the computation time are features and convolving coefficients (decimation).
the common difficulties, and they should be well solved The remainder of this paper has four parts. In section
in all these applications. In texture processing, some au- 2, we give out the function definition of Gabor filter, and
thors consider the design of a single Gabor filter to seg- discuss the parameters selection, and design three different
ment a two-texture image. The output of a Gabor-filtered Gabor filter families. In section 3, four kinds of features ex-
texture is modeled well by a Rician distribution, and a mea- tracted using a Gabor filter family are presented. In section
sure of total output power is used to select the center fre- 4, we do three experiments: experiment A shows the fea-
quency of filter and to estimate the Rician statistics of the tures on the images consisting of different frequency con-
Gabor- filtered image [3]. Other authors utilize the concept tents, experiments B and C show the features on texture and
of multi-resolution with parameter selection to do texture character images. In the last section, we give an overview
analysis, and reduce the computation time by diminishing to the features and the Gabor filter family.
the image size according to several different sets of param-
转载
444-076 139
中国科技论文在线 http://www.paper.edu.cn

2 Designs for Gabor Filter Family

2.1 Gabor Filter Family

The general functional form of a 2-D Gabor filter family
can be specified in eq.(1) and eq.(2), in terms of spatial do-
main impulse response and its frequency domain response:

h(x, y; f, θ) = Figure 1. The real part(left) and imaginary part(right) of a

√ 1
R2 R22 (1)
πσ1 σ2 exp (− 21 ( σ21 + σ22
)) · exp (i(fx x + fy y)) Gabor filter.
1

where R1 = x cos θ + y sin θ, R2 = −x sin θ + y cos θ,

σ1 = cf1 , σ2 = cf2 , fx = f cos θ, fy = f sin θ, c1 and c2
are two constants.
√
The coefficient πσ1 σ2 can guarantee that the ener-
gies of different Gabor
R R filters in the family are all equivalent
to 1, i.e. khk2 = hh∗ dxdy = 1.

H(u, v; f, θ)
√ (2)
= 2 πσ1 σ2 exp (− 21 (σ12 (S1 − f )2 + σ22 S22 ) Figure 2. The frequency response of a Gabor filter.

where S1 = u cos θ + v sin θ, and S2 = −u sin θ + v cos θ.

Gabor filters are spatial sinusoids localized by a Gaus-
The central frequency of the pass band f can be se-
sian window, and they are orientation and frequency sensi-
lected from the interval [0, π]. A series of frequencies in
tive band pass filters. In eq.(1), x and y are the digital pixel
eq.(5) or eq.(6) is usually used
ordinates in the image. The parameters σ1 and σ2 are the
standard deviations of 2-D Gaussian envelope. The central π
fk = , k = 0, 1, 2, · · · (5)
frequency of the pass band is f , and the spatial orientation k+1
is θ.
If let σ1 = σ2 = σ, then eq.(1) and eq.(2) can be π
fk = , α > 1, k = 0, 1, 2, · · · . (6)
simplified as eq.(3) and eq.(4) αk
The sampling in the frequency interval of the former
h(x, y; f, θ) is denser than the latter in general. And in the following of
2
+y 2 (3)
= √1πσ exp (− x 2σ 2 ) · exp (i(fx x + fy y)) this paper, eq.(6) is used.
The orientations can be distributed uniformly in the
H(u, v; f, θ) interval [0, π], such as
√ σ 2 ((u−fx )2 +(v−fy )2 ) (4)
= 2 πσ exp (− 2 ). mπ
θm = , m = 0, 1, · · · , M − 1. (7)
The complex Gabor filter comprises two components. M
R2 R2
One is the real part: √πσ11 σ2 exp (− 21 ( σ21 + σ22 )) cos(f Generally, authors select M = 4 in Chinese character
1 2
R1 ), and the other is the imaginary part: √ 1 recognition, and this just meet the need of the stroke orien-
πσ1 σ2
R2 R2
tations in Chinese characters. And we select the orientation
exp (− 21 ( σ21 + σ22 )) sin(f R2 ). number M = 8 in this paper.
1 2
The real part is even symmetry, whereas the imag- Design 1: If σ1 = σ2 = σ(i.e. c1 = c2 = c), design an
inary part is odd symmetry. For example, fig.1 shows the optimal Gabor filter family.
real part and imaginary part of a Gabor filter, where f = π3 , First we’ll select the constant c in σ(fk ) = fck , then
θ = 0, and c1 = c2 = π. we’ll choose the α in fk = απk .
The frequency response of another Gabor filter is The frequency response of every Gabor filter with the
shown in fig.2, where f = π3 , θ = π4 , and c1 = c2 = π. parameters fk and θm is a 2-D Gauss function as shown
The frequency center in the (u, v) plane is (fx , fy ), where in fig.2, and the pass band width is σ(f2k ) . In the 2-D
f x = f cos θ = 0.7405 and fy = f sin θ = 0.7405. And frequency domain (u, v) ∈ ([−π, π], [−π, π]), The pass
this indicates that a Gabor filter is a Gaussian filter shifted band projection is a circle, whose center and diameter are
to the position (fx , fy ) in frequency. (fk cos θm , fk sin θm ) and σ(f2k ) . These circles and fre-
quency responses are shown in fig.3, respectively.
2.2 Parameters Selection If fk is fixed and θm (m = 0, 1, · · · , M − 1) is
changed, the centers of these circles make up of a bigger
There are four parameters f , θ, c1 and c2 to be selected. circle with the center (0, 0) and the radius fk . To make

140
中国科技论文在线 http://www.paper.edu.cn

and
1 1
+ ≥ fk−1 − fk . (11)
σ1 (fk−1 ) σ1 (fk )
From eq.(10), we can draw c2 ≤ 2M π , and here we se-
lect c2 = 16
π ≈ 5.0930 which is same to design 1. From
eq.(11), we can draw c1 ≤ 3 and select c1 = 3. Fig.5 and
fig.6 show this Gabor filter family and some of its mem-
Figure 3. The pass bands and frequency responses of a bers.
Gabor filter family, where θm = mπ π
8 , fk = 1.4886k and
5.0930
σ(fk ) = fk .

(a)
mπ
Figure 5. Another Gabor filter family, where θm = 8 ,
fk = 2πk , σ1 (fk ) = f3k and σ2 (fk ) = 5.0930
fk .

(b)

Figure 4. The real parts and imaginary parts of some mem-

bers in the above Gabor filter family: (a)different frequen-
π
cies fk = 1.4886 k , k = 0 ∼ 4, (b)different orientations
mπ
θm = 8 , m = 0 ∼ 7.
(a)

these circles be joined, parameters should satisfy eq.(8).

2 (b)
2M · ≥ 2πfk (8)
σ(fk )
Figure 6. Members with (a)different frequencies fk =
π
If θm is fixed and fk (k = 0, 1, 2, · · ·) is changed, the 2k
, k = 0 ∼ 2, (b)different orientations θm = mπ
8 ,m =
centers of these circles make up of a line. To make these 0 ∼ 7.
circles be joined, parameters should satisfy eq.(9).
1 1 Comparing the Gabor filter family in design-2 with
+ ≥ fk−1 − fk (9)
σ(fk−1 ) σ(fk ) the one in design-1, we can find that the former can use the
fewer number Gabor filters than the latter to cover over the
Thus all these circles are joined together and they can same size area in the 2-D frequency domain.
cover over the frequency domain. From eq.(8), we can Design 3: If fk = 2fk0 , k = 0, 1, 2, · · ·, and f0 = 3π
4 , we
educe c ≤ 2M π . Here M = 8, and we select c = π ≈
16
can obtain c1 = 3 and c2 = 5.0930 by the same method.
c+1
5.0930. From eq.(9), we can educe α ≤ c−1 , c > 1. Sub- Thus we get the third Gabor filter family shown in fig.7.
stituting c with 5.0930, we’ll get α ≤ 1.4886. And we
select α = 1.4886.
Since all parameters have been selected, we show
some members of this Gabor filter family in fig.4. These
samples with different frequencies and orientations indi-
cate that each Gabor filter can be got from another one by
dilating and rotating. All of them have the same ridge and
valley numbers.
Design 2: If fk = 2πk , k = 0, 1, 2, · · ·(i.e. α = 2 in eq.(6)), Figure 7. The third Gabor filter family, where θm = mπ
8 ,
design an optimal Gabor filter family. 3π
fk = 4·2 3
and σ2 (fk ) = 5.0930
k , σ1 (fk ) = f fk .
We only need to select the parameters c1 and c2 . k

Similarly, we can get the following two inequalities:

2 We think the third Gabor filter family is more reason-
2M · ≥ 2πfk , (10) able than the second one. Because the pass bands of Gabor
σ2 (fk )

141
中国科技论文在线 http://www.paper.edu.cn

filters with f0 , f1 , f2 , · · · are [ π2 , π], [ π4 , π2 ], [ π8 , π4 ], · · ·, re- • The orientation in which the local has the maximum
spectively. If the central frequency f is cut down a half, energy can be taken as a feature, too.
the bandwidth will be cut down a half, too. And the coeffi-
cients of the Gabor filter decomposition can be done with a F4 (x, y) = k,
radix-2 decimation. where F3 (x, y; θk ) = max {F3 (x, y; θm )}.
m=0∼2
Only design 3 is used in the rest of this paper. (16)
There are some other features that can be extracted
3 Feature Extraction using the Gabor filter family and not enumerated here. And
we can apply one, or several combination of these features
The Gabor filter family captures the whole frequency spec- in our works.
trum, both amplitude and phase. For Gabor feature extrac-
tion, we convolve the image I with every Gabor filter of the
Gabor filter family at every pixel (x, y) as eq.(12) 4 Experiments

G(x, y; fk , θm ) Experiment A. Features On Different Frequencies

P P (12) We have designed two images that contain different
= x0 y0 I(x − x0, y − y0)h(x0, y0; fk , θm ),
frequency components and extracted their features using
where I(x, y) is the pixel intensity. the Gabor filter family. Fig.8 shows our results.
Thus we have got 3 × 8 = 24 output images
G(fk , θm ), k = 0 ∼ 2, m = 0 ∼ 7, and each image
has the same size to the image I. We assume the image
I has X × Y pixels. The pass bandwidths corresponding
to f0 , f1 and f3 are π2 , π4 and π8 , so G(f0 , θm ), G(f1 , θm )
and G(f2 , θm ) can be done with a radix-2 decimation, a
radix-4 decimation and a radix-8 decimation, respectively.
And the sizes of G(f0 , θm ), G(f1 , θm ) and G(f2 , θm ) are Figure 8. The F2 features of two images with different
reduced to X Y X Y X Y
2 × 2 , 4 × 4 and 8 × 8 . Each new el-
frequency components.
ement of G(fk , θm ) after decimation can be the value in
the corresponding block center, or the average value in the
corresponding block. Here we only extract out F2 features. The two images
Now the following G(fk , θm ) all denotes the output are both 64 × 64 pixels. Features F2 (f0 , θm ), F2 (f1 , θm )
image after decimation. and F2 (f2 , θm ), m = 0 ∼ 7 are 32 × 32, 16 × 16 and
8 × 8 matrixes, respectively. And they are shown in se-
• The phase information of G(fk , θm ) can be taken as quential three layers in fig.8. Comparing these feature im-
a feature, because it contains information about the ages, we can find out two facts: one is that the two images
edge locations and other details in the image I. are isotropy and every feature in the same layer contains
equal energy; the other is that the energy distributions of
F1 (x, y; fk , θm ) = phase(G(x, y; fk , θm )) (13) the two images are different: the top one is mainly in the
bands [ π4 , π2 ] and [ π8 , π4 ], but the bottom one is mainly in the
• The amplitude of G(f k, θm ) can be taken as a feature, bands [ π2 , π] and [ π4 , π2 ].
and it contains some oriented frequency spectrum in Experiment B. Texture Feature
every local of the image I. In this experiment, we use two texture images, and
extract out their F2 features like experiment A. Fig.9 shows
F2 (x, y; fk , θm ) = |G(x, y; fk , θm )| (14) the two images and their features.

• The square sum of the different frequency responses

with the same orientation can be taken as a feature,
and it denotes the local energy in certain orientation.

F3 (x, y; θm ) = F2 (x, y; f0 , θm )2
+ 41 F2 (b x2 c, b y2 c; f1 , θm )2 (15)
1
+ 16 F2 (b x4 c, b y4 c; f2 , θm )2 Figure 9. The F2 features of two texture images.
In eq.(15), F2 (f0 , θm ) is a X Y
2 × 2 image, and
X Y
F2 (f1 , θm ) is a 4 × 4 image, and F2 (f2 , θm ) is a These features indicate the energy distribution of ev-
X Y X Y
8 × 8 image. The result F3 (θm ) is a 2 × 2 image, ery image: the top image has a very strong energy at 0 and
π
so each pixel of F2 (f1 , θm ) is split into 4 pixels and 2 orientations; the bottom has about equal energy at every
each pixel of F2 (f2 , θm ) is split into 16 pixels equally. orientation.

142
中国科技论文在线 http://www.paper.edu.cn

If we want to classify or recognize these textures, we

can add up all the square values in every feature image, and
P let all the sums2make up of a 3×8 vector: V [8k+m] =
then
F2 (x, y; fk , θm ) , where k = 0 ∼ 2, m = 0 ∼ 8.
x,y
The vectors produced by the two textures are very un-
like, and they can distinguish the two textures very well Figure 12. The F2 (f2 , θm ) and F4 (f2 ) features of two char-
(shown in fig.10). acter images.

The F4 (f2 ) feature is an 8 × 8 matrix, because it

comes from eight F3 features on the same image. In the
feature image, the strongest brightness denotes the orien-
tation 7π
8 (k = 7), and the weakest brightness denotes the
orientation 0(k = 0). This feature contains the information
Figure 10. The normalized 8 × 3 vectors produced by the of the stroke orientations in the character image.
two texture images.

5 Conclusions
If we want to segment out these textures from other
A Gabor filter family is very similar to a wavelet family,
images, we can first select out some Gabor filters corre-
and the most difference between them is the mode of de-
sponding to those stronger features from the Gabor filter
composing the 2-D frequency domain: the former is circle
family, then convolve the image with the selected Gabor
mode, but the latter is square mode. And thus Gabor fil-
filters, and then add up all the outputs of the Gabor filters,
ters have the orientation characteristics. We have designed
thus the areas that contain the target texture will have the
three Gabor filter families and discussed the differences in
strongest responses in the image, finally we can segment
them, and we commend the third one to users. Using Ga-
out the texture areas from the output easily. For example,
bor filter family, users can extract out many features in the
we can find that the elements V[4] and V[16] in the left vec-
image, such as feature F1 , F2 , F3 and F4 .
tor are much larger than those in the right vector in fig.10,
The three experiments above demonstrate that these
so we use the two relevant Gabor filters to segment out
features extracted using the Gabor filter family well repre-
one texture from the mixed-texture image. Ouput(x, y) =
sent the contents in the image. And these features can be
F2 (x, y; f0 , θ4 )2 + 161
F2 (b x4 c, b y4 c; f2 , θ0 )2 .
used in our practical applications. The parameters of the
For illuminating the performance of this texture seg-
Gabor filter family are preset and are not necessarily opti-
mentation method, we show the results by using one Gabor
mal for a particular task.
filter (F2 (f2 , θ0 )) and multi Gabor filters (F2 (f0 , θ4 ) and
F2 (f2 , θ0 )) comparatively in fig.11. The latter result is bet-
ter. References
[1] John G. Daugman, Complete Discrete 2-D Gabor
Transforms by Neural Networks for Image Analysis
and Compression, IEEE Trans. on Acoustics, Speech
and Signal Processing, 36(7), 1988, 1169-1179.

[2] K.R. Namuduri, R. Mehrotra, and N. Ranganathan,

Edge detection models based on Gabor filters, IEEE
11th IAPR International Conf. On Speech and Signal
Analysis Proceedings, 1992, 729-732.
Figure 11. Texture segmentations using one or multi Gabor
filters. [3] Thomas P. Weldon, William E. Higgins and Dennis F.
Dunn, Efficient Gabor filter design for texture segmen-
tation, Pattern Recogniton, 29(12), 1996, 2005-2015.
Experiment C. Character Feature [4] Chien-Chang Chen and Daniel C.Chen, Multi-
In the last experiment, we give out an example on resolution Gabor filter in texture analysis, Pattern
character feature extraction. The following fig.11 shows Recognition Letters, 17(10), 1996, 1069-1076.
two character images (letters ’0’ and ’A’, size 64 × 64) and
their F2 (f2 , θm ) and F4 features. [5] Jianwei Yang, Lifeng Liu, Tianzi Jiang and Yong Fan,
The F2 (f2 , θm ) features are all 8 × 8 matrixes. A modified Gabor filter design method for finger-

143
中国科技论文在线 http://www.paper.edu.cn

print image enhancement, Pattern Recognition Letters,

24(12), 2003, 1805-1817.

[6] Jouko Lampinen and Erkki Oja, Distortion Tolerant

Pattern Recognition Based on Self-Organizing Feature
Extraction, IEEE Trans. on Neural Networks, 6(3),
1995, 539-547.
[7] V. Tavsanoglu and E. Saatci, Feature Extraction for
Character Recognition Using Gabor-type Filters Im-
plemented by Cellular Neural Networks, 6th Proc.
IEEE International Workshop on CNN and Their Ap-
plication, 2000, 63-68.
[8] Su Yih-Ming, and Wang Jhing-Fa, A novel stroke ex-
traction method for Chinese characters using Gabor fil-
ters, Pattern Recognition, 36(3), 2003, 635-647.
[9] Chih-Jen Lee, Sheng-De Wang, and Kuo-Ping Wu,
Fingerprint Recognition Using Principal Gabor Basis
Function, Proc. of International Symposium on Intelli-
gent Multimedia, Video and Speech Proceeding, 2001,
393-396.

[10] Chengjun Liu and Harry Wechsler, Independent

Component Analysis of Gabor Features for Face
Recognition, IEEE Trans. on Neural Networks, 14(4),
2003, 919-928.
[11] A. Bodnarova, M. Bennamoun and S. Latham, Op-
timal Gabor filters for textile flaw detection, Pattern
Recognition, 35(12), 2002, 2973-2991.

144