----Eg.6 
-——-Edq.7 
3% . 7K) 
——— EQ. 8 
in the IFOV 
) 
dcovers in a 
ng A, which 
liton which 
1e method to 
the | Maximum 
oy Matsumoto 
MULTISCALE APPROACH TO IMAGE TEXTURE 
EA RAE EMO... 
Zhang Jixian 
Institute of Image Recognition & Artificial Intelligence 
Huazhong University of Science and Technology 
Wuhan Hubei 430074 
P. R CHINA 
‘Commission II, Working Group 2 
KEY WORDS: Texture, Classification, Feature, Fusion, Extraction, Texture Analysis, Gabor Function, 
Multiscale Decomposition 
ABSTRACT: 
It is important to consider the role of scale for texture analysis since its multiscale attribute of image tex- 
ture. In this paper, a textural detector based on 2D Gabor function and visual textural perception is estab- 
lished first, then based on the textural detector and wavelet theory of multiscale decomposition and fractal 
geometry, a multiscale texture analysis method is proposed, and technique for multiscale textural feature 
fusion is advanced according to the lateral inhibition and end-inhibition in neurodynamics. The mult- 
iscale texture analysis technique gives representation between spatial space and Fourier space, and pro- 
vide a hierarchical analysis framework for image texture. They can detect different scale texture features, 
correspond to the visual texture perception, and have the ability to recognize texture image effectively. 
1. INTRODUCTION 
Image texture analysis has become fundamental 
means in the areas of computer vision and image 
analysis. So far many methods have been developed 
for the description of textural features (Deren Li 
and Jixian Zhang,1993), however, most of them. 
extract textural features only in some one scale and 
ignore its multiscale attribute of image texture, 
general-purpose, universally accepted method is still 
unavailable. 
Inspired by a multi-channel filtering theory for 
processing visual information in the early stages of 
the human visual system, multi-channel filtering ap- 
proach to texture analysis is developed, however 
following issues are unsolved: (1) mathematical 
functional indication and the number of multi- 
channel filters; (2) detection of suitable texture fea- 
tures and integration among these features in 
filtered images; (3) relationship among filtered 
images. 
According to our proposed methodology (Jixian 
Zhang,1994), image texture is regarded as the 
spatial distribution of grey levels of neighboring 
pixels, it has hierarchical attribute, multiscale 
attribute, shift-invariant attribute and stochastical 
and deterministic duality. Image texture analysis 
method should existed in a hierarchical framework, 
while extraction of image texture feature should 
consider its multiscale attribute. In this paper, a 
textural detector based on 2D Gabor function and 
visual textural perception is established first, then 
based on the textural detector and wavelet theory, a 
multiscale texture analysis method is proposed, and 
technique for multiscale texture feature fusion is 
advanced, finally some experiments are gived. 
2. MODEL OF VISUAL TEXTURAL | 
DETECTOR 
According to the  preattentive theory, visual 
discrimination of image texture is achieved by two 
steps: (1) detection of local feature difference---- 
texton (or textel); (2) discrimination based on 
statistical feature of detected textons(Julesz 1986). 
It is important to find the function of textural 
detector for image texture analysis, which should 
not only has the ability to detect any kinds of textels 
effectively, but also correspond to the visual texture 
perception. 
Two-dimensional (2D) Gabor representation gives 
an attractive framework for a unified theory and 
mathematical description of the spatial receptive 
fields of visual cortex (Daugman 1988), such filters 
simultaneously capture all the fundamental proper- 
ties of linear neural receptive fields in the visual 
cortex: spatial localization, spatial frequency 
. selectivity, and orientation selectivity. Any image 
can be expanded by a finite set of 2D elementary 
Gabor functions and the expansion coefficients {a} 
provide a compact representation of the image. 
Experiments by Fogel and Sagi (1989) showed that, 
by using 2D Gabor filters, results to discriminate 
textural elements used in Krose's psychophysical 
data are in high correlation with the results for the 
999 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996