t OBJECT 
)N 
ested objects in 
inal color aerial 
' Fuzzy C-Mean 
s presented that 
is is because the 
influence of the 
ation of photo 
'ed can improve 
/ith realistic and 
n aerial images 
righest building 
:xtract the huge 
e the parallaxes 
"the image and, 
ine and produce 
dels are created 
Virtuozo), the 
> that parallaxes 
d into different 
gorithm. Third, 
e original color 
id low objects. 
objects mainly 
and more non- 
T to refine trees 
ed by Fuzzy C- 
this paper, an 
n forest and on 
bs in automatic 
acy of observed 
ts on non-trees. 
erved points on 
ps are given the 
eriment results 
its divided into 
y of automatic 
2 describes the 
high and color 
on. In section 3, 
Jlusion is given 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
2. The objects extraction based high and color texture 
features 
In this session, we describe our segmentation algorithms, 
including preliminary segmentation based on disparity image, 
the texture features and final trees of extraction using fuzzy c- 
mean. 
2.1 High and low objects distinguishing 
We start with DEM data an automatically generated by the 
digital photogrammetry system --Virtuozo. The resolution of 
the images by which DEM data are obtained may be lower than 
original images. According to following algorithm high and 
low objects are obtained. 
Algorithm 1: Preliminary segmentation based on high features 
Given: original color aerial images 
Step 1. DEM are mapped to range (0—255 gray level) in order 
to form the image of DEM. As a result, different gray 
levels denote different elevation. 
Step 2. With the DEM image, original image is divided into 
many regions with same size . 
Step 3. The edges in the DEM image are extracted by Sobel 
algorithm. These edges reflect the local changes of 
elevations of objects. 
Step 4. According to the edge image, we compute the segment 
threshold for each region according to following rules: 
The next step is to refine trees from high objects by Fuzzy C- 
Mean clustering based the color texture features. 
2.2 The classification based on Fuzzy c-mean 
Fuzzy c-mean clustering algorithm (FCM) was introduced 
by J. C. Bezdek (J. C. Bezdek, 1987). In this paper, FCM is 
used to refine the trees from the high and low objects. The 
algorithm is described as following. 
Algorithm 2: The classification based on Fuzzy c-means 
clustering. 
Given : The images including the high and low objects. 
Step 1: Calculating the texture and color features values of 
every pixel in the images(. 
Step 2 : Computing the fuzzy membership of the pixels of 
images. 
e n 2 à; 
Un "5 POULE 
JR 
: Un" X, 
y vi (1) 
! n m 
kzl Us 
  
Disparity data 
  
I 
  
  
Disparity image 
  
  
RE 
  
Edge detection and threshold 
  
iL 
il 
  
Low objects 
  
  
  
High objects 
  
  
  
  
  
Fuzzy C-Means 
  
  
  
Raw images 
(1) If there is any edge in a region, the average gray 
level value of pixels on edges is to set to the 
threshold; 
(2) If there is no edge in a region, the threshold can be 
obtained by the bilinear interpolation method of 
thresholds of its neighbor regions. 
Step 5. According to the threshold of region, the original image 
is segmented into a binary image, in which | 
represents high objects and 0 represents low objects. 
The high objects include trees, houses ,bridges and so 
on. 
  
Figurel. The algorithm procedure 
where D;, is some measure of similarity between v; and x, or 
the attribute vectors, and the cluster centre of each region 
v 2 (v, v3, K ,v,) is geometric cluster prototypes. U denotes 
the fuzzy membership matrix of pixel block k in cluster i, c 
denotes the number of cluster. 
Ussiscestsn (2)