You are on page 1of 4

Double-Density Discrete Wavelet Transform Based Texture Classification

Yu-Long Qiao
1
, Chun-Yan Song
2
, Chun-Hui Zhao
1
1
College of Information and Communications Engineering,
Harbin Engineering University, Harbin, China
2
College of Information and Computer Engineering,
Northeast Forestry University, Harbin, China
qiaoyulong@hrbeu.edu.cn; yulongqiao@hotmail.com
Abstract
Texture classification plays an important role in
image analysis. The wavelet is a very efficient
multiscale analysis method that has been successfully
applied to describe the texture. However, it is
translation-invariant. The recent double-density
discrete wavelet transform have two interesting
property, low computational complexity and nearly
shift invariant. In this paper, the texture feature based
on the double-density discrete wavelet transform are
derived from the subbands and tested in two texture
databases. The experimental results suggest the
potential capacity of the double-density discrete
wavelet transform in texture analysis.
1. Introduction
Texture classification plays an important role in
image analysis and understanding. It has been broadly
applied to remote sensing, automated inspection,
medical image analysis and document processing. In
addition, many recent content-based image retrieval
methods take the texture as a key feature. Much crucial
work has been done in recent years. Despite the effort,
it is still an interesting and challenging research field.
There are some reports [1, 2] surveyed on texture
analysis methods. Tuceryan and Jain [2] identified four
major categories of methods to extract textual features:
statistical, geometrical, model-based, and signal
processing based methods. Initially, texture analysis
was based on the first-order and second-order statistics,
in which the co-occurrence matrix [1] is one of the
most important methods. Afterwards Markov random
field [3], fractal [4], wold decomposition [5] and
various transforms were successfully applied to this
field. However, a common weakness of the early
methods is that the texture is analyzed only at a single
scale.
Law [6] convoluted some designed masks with a
texture and calculated energy statistics to describe the
texture. His work introduced the concept of
multichannel processing. In the field of signal
processing, there are some interesting
spatial/frequency analysis techniques such as the
Gabor transform, the Wigner distribution and the
wavelet transform provide multiresolution analytical
tools.
Since Mallat [7] presented the wavelet
representation and showed it can be used in many
aspects of image processing and analysis. Many
famous techniques for image analysis based on the
wavelet have been proposed. Although the discrete
wavelet transform possesses many useful
characteristics and achieves satisfying results in many
fields such as image compression, it is translation-
variant. That is to say, when an image pattern is
translated, its numerical descriptors will be not
translated but modified. This behavior is inadequate in
texture analysis, for one usually thinks of texture as
a translation-invariant property. One solution is to use
the overcomplete transform.
The double-density discrete wavelet transform
proposed by Salasnick [8] has two interesting property,
low computational complexity and nearly shift
invariant. In this paper, this transform is applied to
texture classification. The paper is organized as
follows. The next section provides an introduction to
the double-density discrete wavelet transform. In
Section 3, the texture classification method is
presented. The experimental results are showed in
Section 4.
2. Double-Density DWT
The double-density DWT is characterized by one
scaling function ) (t | and two distinct wavelets ) (
1
t
and ) (
2
t . Following the multiresolution framework,
they should satisfy the dilation and wavelet equations,
2 , 1 ) 2 ( ) ( 2 ) (
) 2 ( ) ( 2 ) (
0
= =
=
_
_
i n t n h t
n t n h t
n
i i
n
|
| |
where 2 , 1 , 0 ), ( = i n h
i
are filters and satisfy the
perfect reconstruction condition
0 ) 1 ( ) (
) 1 ( ) ( ) 1 ( ) (
2 ) 1 ( ) (
) 1 ( ) ( ) 1 ( ) (
2 2
1 1 0 0
2 2
1 1 0 0
= +
+
= +
+
z H z H
z H z H z H z H
z H z H
z H z H z H z H
Let ) ( k t
k
= | | , ) 2 ( ) (
1 , , 1
k t t
j
k j
= and
) 2 ( ) (
2 , , 2
k t t
j
k j
= , then any square integrable
function f(t) can be represented as
___ _
=

=
+ =
2
1 0
, ,
) ( ) , ( ) ( ) ( ) (
i j k
k j i i
k
k
t k j d t k c t f |
The two distinct wavelets ) (
1
t and ) (
2
t are
specifically designed to be offset from one another by
one half such that ) 5 . 0 ( ) (
2 1
~ t t in [8]. It
creates an approximately shift invariant transform.
To perform the double-density discrete wavelet
transform on an image, we should alternatively apply
the transform first to the rows, then to the columns of
the resulting images. This gives rise to nine 2-D
subbands, one of which is the 2-D lowpass scaling
filter, and the other eight of which make up the eight 2-
D wavelet filters, as shown in Figure 1.
Figure 1. 2D double-density DWT
3. Texture Classification
Texture classification consists of two phases,
learning and classification. In the learning phase, the
texture feature is extracted and the classifier is
designed. In the second phase, the unknown texture
image is classified through the designed classifier after
the feature extraction.
3.1. Feature Extraction
Given a texture image, there are L 8 detail
subbands after L-level double-density wavelet
transform. It is assumed the texture images are
spatially homogeneous, and the mean and the standard
deviation of the magnitude of the transform coefficient
of these detail subbands
__
= =
=
M
m
N
n
i i
n m WC
MN
1 1
) . (
1

( )
__
= =
=
M
m
N
n
i i i
n m WC
MN
1 1
2
) , (
1
o
can be used to describe the texture image, where
WC
i
(m,n) and N M are the coefficient and size of
the ith subband, respectively. The feature vector of a
texture image, which is referred as TF1, is now
constructed by gathering the mean and the standard
deviation of each detail subband.
When features based on first-order statistics do not
suffice, the second-order statistics should be
considered to improve the texture classification.
Therefore, in order to improve the classification
performance, the co-occurrence matrix feature is also
included in the feature vector.
Due to the computational complexity, the co-
occurrence matrix should be computed on an image
with a reasonable gray level. However, the wavelet
coefficients are real number. Hence we need to
quantize these coefficients. The method is to choose
2J+1 intervals J J j u u
j j
, , 0 , , ), , [
1
=
+
, and
then let
) , [ ) , ( , ) , (
1 +
e + =
j j i i
u u n m WC if J j n m WC
The co-occurrence matrices are formed for each
quantized detail subband for the horizontal direction
and the vertical direction with the unit distance using
the definition of the co-occurrence matrix. From this
matrix, two features
_
_
__
__
=
=
+ =
=
) , (
) , ( where
) , ( ) ( 2
) , ( ) ( 1
2
2
j i jC c
j i iC c
j i C c j c i c
j i C j i c
y
x
y x


are extracted, where C is a co-occurrence matrix. For
convenience, the dependence on the subband label and
0
h
1
h
2
h
2 !
2 !
2 !
0
h 2 !
2 !
2 !
1
h
2
h
0
h 2 !
2 !
2 !
1
h
2
h
0
h 2 !
2 !
2 !
1
h
2
h
the direction is suppressed. The resulting feature vector
is denoted by TF2.
3.2. Classification
In the classification phase, an unknown image is
decomposed using the double-density discrete wavelet
transform and its features are calculated in the same
manner as the learning phase to form the feature vector.
Then, the k-nearest neighbor classifier is used to
classify the unknown image with the normalized city-
block distance
_
=
n
n n i n u
f f i D o
, ,
) (
where
n u
f
,
and
n i
f
,
are the feature vector of the
unknown image and the feature vector of ith texture
image in the database, respectively,
n
o is the standard
deviation of the respective features over all the training
samples.
4. Experimental results
Our experiments on texture classification are
conducted with two benchmark texture databases. The
first one contains 1920 image samples with 30 texture
classes chosen from the Brodatz texture database [9],
which is constructed by dividing each 512512 image
into 64 nonoverlapping regions. 32 texture samples are
used to design the classifiers; the others are adopted as
test samples. The other database is generated with the
same method, but the original textures are taken from
the VisTex database [10].
First, the experiment is performed on the 30
Brodatz texture images with two different discrete
wavelet transform (DWT), critically-sampled DWT
(CS DWT) and Double-Density DWT, and two
different features, TF1 and TF1+TF2. The correct
classification number for each texture image is listed in
Table 1. The average classification rate (Avg.) is also
shown in this table. When the double-density DWT is
adopted, the Avg. of TF1 and TF1+TF2 are 91.3% and
93.1%, respectively. For the CS DWT, the average
classification rates are 88.2% and 89.8%. The largest
increment is 4.9%.
The same experiment is done on the Vistex
database and the result is show in Table 2. The average
classification rates of TF1 and TF1+TF2 are 90.4%
and 91.6%, respectively, when the CS DWT is used.
For the double-density DWT, the corresponding
average classification rates are 92.7% and 94.6%.
These classification performance shows that even the
Avg. of TF1 in the double-density DWT domain is
better than that of TF1+TF2 in the CS DWT domain.
5. Conclusions
The classification performance of the double-
density DWT is compared with the critically-sampled
DWT using two classes of features and two different
texture databases. When the double-density DWT is
used, the largest increments are 4.9 % and 4.2%,
respectively. The average classification rates shows
that even the performance of TF1 of the double-density
DWT is also better than that of TF1+TF2 of the CS
DWT. Take the computational complexity and the
classification performance into account, the double-
density discrete wavelet transform is a better choice for
texture classification.
References
[1] R. M. Haralick, Statistical and Structural Approaches to
Texture, Proceedings of IEEE, 67(5) (1979), pp. 786-804.
[2] M. Tuceryan, and A. K. Jain, Texture Analysis, In
Handbook of pattern recognition and computer vision (2ed),
World Scientific Publishing Co., 1998.
[3] G. C. Cross, and A. K. Jain, Markov Random Field
Texture Models, IEEE Trans. Pattern Anal, Machine Intell.,
5(1) (1983), pp. 25-39.
[4] A. P. Pentland, Fractal-based Description of Natural
Scenes, IEEE Trans. Pattern Anal, Machine Intell., 6(6)
(1994), pp. 661-674.
[5] F. Liu, and R. W. Picard, Periodicity, Directionality, and
Randomness: Wold Features for Image Modeling and
Retrieval, IEEE Trans. Pattern Anal, Machine Intell., 18(7)
(1996), pp. 722-733.
[6] K. I. Law, Texture Image Segmentation, Ph.D.
dissertation, Univ. of Southern California, (1980).
[7] S. G. Mallat, A Theory for Multiresolution Signal
Decomposition: the Wavelet Representation, IEEE Trans.
Pattern Anal, Machine Intell., 11(7) (1989), pp. 674-693.
[8] I. W. Selesnick. The double density DWT. In A.
Petrosian and F. G. Meyer, editors, Wavelets in Signal and
Image Analysis: From Theory to Practice. Kluwer, 2001
[9] P. Brodatz, Textures: A Photographic Album for Artists
& Designers, New York: Dover, (1966).
[10] VisTex database, available at http://vismod.media.
mit.edu/vismod/imagery/VisionTexture /vistex.html
Table 1. Experimental results on Brodatz
images
CS DWT
Double-Density
DWT
TF1 TF1+TF2 TF1 TF1+TF2
d005 32 31 32 32
d006 32 32 32 32
d009 32 32 32 32
d014 32 32 32 32
d018 32 32 31 32
d020 32 32 32 32
d022 31 31 30 32
d026 32 32 32 32
d030 14 20 20 28
d034 32 32 32 32
d036 31 30 32 31
d041 30 32 32 32
d045 20 27 19 25
d047 32 32 32 32
d048 27 26 28 26
d052 32 31 32 32
d055 32 32 32 32
d062 22 28 16 24
d064 32 32 32 32
d069 32 31 32 32
d073 28 30 28 32
d075 32 32 30 31
d082 32 32 32 32
d088 12 5 20 15
d091 14 16 13 17
d099 13 15 23 25
d101 32 32 32 32
d104 32 32 32 32
d107 29 29 32 32
d109 32 32 32 32
Avg. 88.2% 89.8% 91.3% 93.1%
Table 2. Experimental results on VisTex
images
DWT
Double-Density
DWT
TF1 TF1+TF2 TF1 TF1+TF2
Bark0 32 32 32 32
Bark4 29 30 28 32
Bark6 32 29 32 28
Bark8 23 23 25 27
Bark9 17 17 18 22
Brick1 31 31 32 31
Brick4 28 28 31 29
Brick5 29 29 32 29
Fabric0 32 32 32 32
Fabric4 31 32 31 29
Fabric7 32 32 32 32
Fabric9 32 32 31 32
Fabric11 21 21 26 31
Fabric13 32 32 30 32
Fabric16 29 28 31 31
Fabric17 32 32 32 32
Fabric18 32 32 32 32
Food0 28 27 30 31
Food2 22 23 23 24
Food5 32 32 32 32
Food8 31 31 30 31
Grass1 27 28 31 28
Sand0 32 32 32 30
Stone4 32 32 32 32
Tile1 25 25 25 32
Tile3 27 30 26 32
Tile7 32 32 32 32
Water6 32 31 32 31
Wood1 22 29 26 28
Wood2 32 32 32 32
Avg. 90.4% 91.6% 92.7% 94.6%

You might also like