84
TEXTURE CLASSIFICATION RESEARCH
BASED ON LIFTING-BASED DWT 9/7 WAVELET
Hong Zhang a ’ *, Ning Shu a
a School of Remote Sensing Information Engineering, Wuhan University, 129 Luoyu Road, Wuhan 430079, China
Email: hongzhangwh@163.com
Modified update Ui
jlSdf+a],
a ‘{ S(d}+dl
Modified update P1
, l a ( X 2i + X 2i
'{2ax 2i +x 2i
KEY WORDS: lifting-based DWT 9/7 wavelet; Wavelet statistical features (WSFs); Wavelet co-occurrence features (WCFs);
Feature extraction; Texture classification
ABSTRACT:
Texture classification analysis plays an important role in remote sensing image. The main difficulty of texture analysis in the past
was the lack of adequate tools to characterize different scales of textures effectively. The 2-D lifting-based DWT 9/7wavelet filter is
used here, without additional computations, giving lifting-based architectures a significant advantage over convolutional filter band-
based architectures. This paper describes the texture classification using (1) the known texture images are decomposed using
9/7wavelet. Then, mean and standard deviation of approximation and detail sub-bands of 3- level decomposed images are calculated.
They are wavelet statistical features (WSFs) (2) In order to improve the correct classification rate further, it is proposed to find co
occurrence matrix features for original image, approximation and detail sub-bands of 1-level 9/7 wavelet decomposed images. The
various co-occurrence features such as contrast, energy, entropy and homogeneity are calculated from the co-occurrence matrix.
These are wavelet co-occurrence features (WCFs). (3) At last, the combination of WSFs and WCFs (feature vector) are used to
classify images. From the exhaustive experiments conducted with texture images, it is found that (1) using lifting-based DWT 9/7
wavelets filter, it is easily adapted to arbitrary remote sensing image sizes, yielding texture classification analysis system that is
small in size, low in memory, and high in performance. (2) At the same time, the mean success rate is improved for a combination of
WSFs and WCFs compared to that of WSFs (or) WCFs.
1. INTRODUCTION
As we all known, texture classification analysis plays an
important role in remote sensing image. The main difficulty of
texture analysis in the past was the lack of adequate tools to
characterize different scales of textures effectively. The 2-D
lifting-based DWT 9/7wavelet filter and Gray Level Co
occurrence Matrix is used here, we can get the curve of from
the average energy of sub-image of used land samples, and
texture classification parameters of some kind of land
samples.Part two give the introduction of CDF 9/7 wavelet,
part three give the selecting texture structure image character
method based on wavelet transform and Gray Level Co
occurrence Matrix ,and part four give the result.
2. INTRODUCTION OF CDF 9/7 WAVELET
Lifting scheme can design the wavelet transforms that map
integer to integer, which provides the effective tool for studying
the lossless image compression. The method for wavelet
coefficients coding is a key technique to implement image
compression, which not only affects the effect of compression,
but also the qualities of resuming image and the time of coding
and decoding. It not only inherits the characteristic of the first
generation wavelet, but is independent of Fourier transform and
particularly easy to build on non linear wavelet transforms, such
as integer wavelet transforms.
The Cohen-Daubechies-Feauveau 9/7 (CDF 9/7) wavelet
transform is an especially effective biorthogonal wavelet, used
by the FBI for fingerprint compression and selected for the
JPEG2000 standard. CDF 9/7 wavelet have good property,
using lifting scheme method, map integer to integer is come
true.
The biorthogonal 9/7 wavelet can be implemented as four
lifting steps followed by scaling; a lifting scheme requires that
the following equations be implemented in hardware:
predict PI: d) = a(x 2i + x 2i+2 ) +x 2i+x
update U1: a ) = (3{d) + d)_ x ) + x 2i
predict P2: df = y(a\ + a) +x ) + d)
update U2: a f = S(df + d*_ x ) + a]
Scaled: a i =gaf
Scale G2: d, = d] / g
The original data to be filters is denoted by Xi, and the 2-D
DWT outputs are the approximation coefficients and details
coefficients.
When the biorthogonal 9/7 wavelet is implemented with
convoltional filter, the image is extended by four pixel on all
sides, the simple way of implementing symmetric extension is
used, and the update step equations are modified to change the
calculation for the first pixel in each row and column so that
Modified update U1:
fi(d\ + d)_ x ) + x 2l , i = 1,2,..., N / 2-1
* Corresponding author.
Modified update P2
2 _ Y( a ) +a '>l
'{2ya]+d
There are some
0.05298011854, 1
4 =1.149604398.
L*L image 3 re
follows:
image f(x,y
(a) original im
112
HL2
LH2
HH2
LH1
\
(c) second level dec
Figure 1. 2-d DWT
Thereinto: LL: app
HL: plai
LH: ven
HH: dia
3. EXTRACT]
FEATURE BAS1
GRAY let
Generally, the resu
character measure
wavelet transform
texture is character
a knighthood, the 1
is repeated over the
°r texel. Selecting
wavelet transform
useful.
