(2010) presented a method of medical image texture features 
classification by using Gabor wavelet model. This paper will 
utilize Gabor filters to extract texture features of multispectral 
and SAR images. 
As few people are working on researching the transformational 
methods of texture features description under different imaging 
conditions so far, few papers can be found in this area. And this 
will be the main research content of our paper. 
2. TEXTURE DISCRIPTION BASED ON GABOR 
WAVELET 
Research has demonstrated that Gabor filter is the optical one 
when describes texture both in spatial domain and frequency 
domain. Its orientation, bandwidth and center frequency can be 
changed according to different requirements. People design 
optical Gabor filters for texture features description based on 
the theory that texture belongs to the narrowband signal and 
different textures always have different center frequency and 
bandwidth when it transforms to the frequency domain from 
space. Each filter will show one kind whose orientation and 
frequency is same as the filter’s only. People will get a set of 
image texture descriptions after using Gabor filters for image 
filtering. 
Actually, the Gabor wavelet can be considered as a special 
wavelet transform. An image’s 2-d wavelet transform is 
F(x) = [FG px, y)dxdy (0) 
where f(x,y)7 the value of image gray 
g(x, y) 7 the mother wavelet 
The Gabor filtered output of an image f(x,y) is obtained 
when the mother wavelet g(x, y) in Eq.(1) is replaced by the 
Gabor function g(x, y), given in Eq.(2) 
1 (x 2 
g(x,y) =| —— exp -=| 5+ 4 
220,0, 26; 6, 
The Eq.(2) shows that a 2-d Gabor function is Gaussian 
modulated sinusoid. The parameters of a Gabor function are the 
  
  
Jess. 6 
modulation frequency Jy , the orientation Q and the Gaussian 
function's scale c, and c. (Li, Meng, 2008). 
In frequency domain, Gabor wavelet can be obtained by 
moving Gaussian function along y -axis. The corresponding 
representation in frequency domain is 
G(u,v) = a a (3) 
H v 
    
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
  
   
The g: and g.. in Eq. (3) are the radius of 4/-axis and V - 
M s q H 
axis in frequency domain. They can be calculated as: 
ide 321 (4) 
  
A set of self-similar filters can be generated from the dilation 
and rotation of the mother wavelet. And the rotated equation is 
g,(y)2a"g(x, y): a»1 (5) 
a" = the scale factor 
x'=a”"(xcos0 + ysin 0) 
where 
y'=a”"(=xcos0 + ysin 0) 
k = the total number of orientations 
A. set of Gabor filters are convolved with the image shown in 
figure.l, and the outputs for six orientations are shown in 
figure.2. 
  
AN 
(e) 
Figure 2. Textures of Lena with different 
orientations(a to f stand the texture images of g — 7A to g2z 
with every 7 
ry A )