FEATURE FUSION BASED ON DEMPSTER-SHAFER'S EVIDENTIAL REASONING 
FOR I MAGE TEXTURE CLASSIFICATION!* 
“Jia Yonghong , "Li Deren 
“School of Remote Sensing Information Engineering, Wuhan University, Wuhan, China, yhjia2000@sina.com 
"LIESMARS, Wuhan University, Wuhan, China, dli@wtusm.edu.cn 
Commission III, WG III/8 
KEY WORDS: Feature fusion, Dempster-Shafer's evidential reasoning, Fractal, Grey co-occurrence matrix, Image texture 
ABSTRACT: 
A new multi-feature fusion technique based on Dempster-Shafer's evidential reasoning for classification of image texture is 
presented. The proposed technique is divided into three main steps. In the first step, the fractal dimension and gray co-occurrence 
matrix entropy are extracted from a texture image. In the second step, we focus on how to design a probability assignment function 
m(A) representing the exact belief in the proposition 4 depicted by one of features. A combining rule, which synthesizes 
probability assignment functions representing the fused information, is proposed based on Dempster-Shafer's evidential reasoning. 
The formulas for calculating the belief function Belief(A), the plausibility function Plausibility (4) and uncertainty probability are 
given. In the decisive step in which image texture is classified, a set of decision rules is provided. An example is provided, and the 
performance is investigated with some aerial photos. Texture classification is considered, with the following classes: inhabitant area, 
water field, grassland and woodland. As a reference for evaluating the performance of multi-feature fusion technique based on 
Dempster-Shafer's evidential reasoning in texture classification, classification accuracies using the single-feature and fused features 
are calculated. Compared with the results obtained from the single feature, the results obtained from multi-feature fusion indicate the 
multi-feature fusion technique based on Dempster-Shafer's evidential reasoning for classification is stable and reliable, and 
efficiently improves the accuracy of classification. 
1. INTRODUCTION 
The aims of texture analysis are texture recognition and texture- 
based shape analysis. A variety of statistical methods such as 
primitive length features [Galloway M., 1975], edge frequency 
method, autocorrelation [Haralick R. 1979], co-occurrence 
approach [Haralick R. 1986], fractal dimension and Markov 
random field method [Huang Guilan, Zheng Zhaobao, 1998a, 
1998b], etc., have been proposed for texture analysis which are 
based on capturing the variability in gray scale images. One of 
the best methods is that parameters of Markov random field 
model, features of gray scale co-occurrence matrix and fractal 
dimensions extracted from image are combined, and then fuzzy 
clustering analysis are applied for image texture classification 
[Huang Guilan, Zheng Zhaobao, 1998a, 1998b]. But it has two 
shortcomings, the one is its complexity and incompleteness in 
obtaining parameters of Markov random field model by 
Bayesian decision; the other is not concerned features whether 
are related or are the best combined. Extraction and selection of 
image texture features in classification are very important, 
classification according to only one feature has its localization 
in accuracy, and can't satisfy the requirement of identifying 
image targets. So a multi-feature fusion technique based on 
Dempster-Shafer's evidential reasoning for image texture 
classification is presented. The remainder of this paper is 
organized as follows. The methodology is explained in section 
2. The experiments and discussion are given in section 3. 
Finally, the conclusions are summed up in section 4. 
  
2. FEATURE FUSION BY DEMPSTER-SHAFER'S 
EVIDENTIAL REASONING FOR I MAGE TEXTURE 
CLASSIFICATION 
2.1 Measurement of fractal dimensional feature 
A variety of methods measuring fractal dimension from image 
texture have been proposed. Here is given a method of 
measuring Brown's fractal dimension. 
Supposing XE" (E" is a n-dimensional space), AX) is a real 
random function. If a constant H(0<H<I, called self-similar 
parameter) exist, function F(t) is a distributing function having 
nothing with X or L.X, then f(X) is called Brown's fractal 
function. Its expression is 
JOX* AX )- f(X) 
Fly=d, (omy) (1) 
And its fractal dimension is 
f=n+l-H (2) 
Eq. (1) can be rewritten as 
"The project supported by the National Surveying and Mapping Fund of China 
662 
   
   
  
  
  
    
   
   
  
  
  
  
    
  
  
  
   
   
   
   
   
   
   
  
   
  
  
  
   
   
  
    
  
   
   
  
   
  
  
   
   
    
  
Internat 
Therefc 
Where 
Therefoi 
function 
texture, | 
where 
yi 
The frac 
First of 
respectiv 
y) fally 
and /gC 
accordin 
accordin 
2.2 Me: 
Occurrei 
The co- 
image f 
J[0:g-1] 
direction 
P(ij, 
Separate 
combina 
based on 
Once the 
can be 
invariant 
distancec 
all direct 
are comp 
features 
Occurren 
character 
Experim: 
matrix i: 
power. H