International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
2.3.2 Designing for the basic assignment function: Designing 
for the basic assignment function m,(/) is: first, we set up a 
model base of texture image by selecting a variety of texture 
training samples according to prior knowledge; then we extract 
features f; (i=1,2) from unknown texture, and match them with 
features of image texture in model base respectively to obtain 
correlation coefficients Pj), here / represents the class number 
of texture in the model base; finally m£j) which feature f; 
assigns texture j can be constructed by Pj) . Its expression 
[Yang Jingyu, Wu Yongge, Liu Leijian Etc. 1994] is given as 
follow: 
yn Sls 
m am Xen d 
J 
Where 
a; = maxP;(j) is the maximum correlation coefficient 
J 
between feature /; of unknown texture and that of texture j 
  
Toh i D St fficient of feature f, 
B, X^ is distributing coefficient of feature f; 
/ 
Vi = SI is reliable coefficient of feature /;.. 
Therefore 
nz ANUS... (13) 
Y p, Gn(0o.- y; X1-a;B;) 
J 
Here m; (6) is an uncertainty probability which feature f, 
assigns frame of discernment 6. 
According to Eq. (12), we can compute the single feature 
(fractal dimension or entropy) probability mass function, 
then fuse according to Eq.(11) probability mass functions of 
2.3.3 Decision rules for texture classification: Based on the 
analysis of the combined probability mass function, we adopt 
four decision rules to classify textures: 
*The target should have the largest probability mass function; 
*The difference between the target probability mass function 
and any other target's should be greater than threshold 4, 
namely that the supporting degrees of all textures by each 
possible proposition should have difference enough. 
*The uncertainty probability m6) must be smaller than 
threshold /2. 
*The target probability mass function must be greater than 
m0). When We seldom know a target, the classification can't 
be preceded. 
Above all, the processes of image texture classification based 
on the Dempster— Shafer reasoning theory are given as follows: 
First of all, m{A),Belief{A), and Plausibility(A) of each feature 
are computed; Secondly, according to the combining rule, 
compute the fused probability mass function m; and its Belief; 
and Plausibility;, Finally, according to decision rules to choose 
the maximum hypotheses in the action of fusion. 
3. EXPERIMENT AND DISCUSSION 
The performance of the method is investigated with some aerial 
photos on some area. A Four-class texture classification is 
considered, with the following classes: inhabitant area, water 
field, woodland and grassland. 10 test samples from each 
texture are obtained; the size of the test sample is 100 by 100 
pixels. As a reference for evaluating the performance of multi- 
feature fusion technique based on Dempster-Shafer's evidential 
reasoning, the same decision rules (/,70.1,/570.3) are utilized, 
classification accuracies with the  single-feature (fractal 
dimension or entropy of co-occurrence matrix) and the fused 
feature are calculated. See Table 1. Compared with the results 
obtained from the single feature, the results obtained from 
multi-feature fusion have higher accuracy of classification, 
single feature, either the fractal dimension or entropy of gray 
co-occurrence matrix, is not sufficient for descripting texture. 
This indicates the multi-feature fusion technique based on 
Dempster-Shafer's evidential reasoning for classification is 
  
  
all features to get the combined probability mass function. stable and 
reliable, and efficiently improves the accuracy of image texture 
classification. 
Feature Fracta Entropy of grey Multi- 
I co-occurrence feature 
featur matrix fused 
e 
Accuracy of | 72.5% 80% 95% 
classification 
  
  
  
  
  
  
Table 1. The accuracy of texture classification based on Dempster-Shafer evidential reasoning. 
664 
Intern 
Table 
probat 
feature 
corres] 
feature 
corres] 
Demps 
to ide 
probab 
enhanc 
A new 
Shafer' 
is presc 
steps. / 
is inves 
with th 
obtaine 
fusion 
reasonis 
improve 
Gallow: 
run len 
pp172-1 
Haralic 
texture. 
Huang 
fractal 
e 
Survey 
Huang ( 
in imag 
No.1, pp 
Li Derer 
texture a