MXM. Here M is the size of the image blocks which are 
divided in the following steps. It is because that texture feature 
is a property of area. Another supposition is that any texture 
area in image should include at least one foundational element 
of the texture(Zhou, 2001). 
At the first step, the initial segmentation image will be gotten 
by the following process. First, the original image is divided 
into many sub-block with size of M><M. Texture feature of 
every sub-block is calculated by the equations above. Second, a 
feasible clustering algorithm is chosen to classify them to 
certain amount classes. And different classes are marked by 
different gray level. Edges between different areas in initial 
segmentation image are ladderlike. In order to get smooth and 
rational edge, it is necessary to perform the following edge 
fining algorithm. The algorithm, initial value and parameters of 
clustering should be confirmed in this step. 
At the second step, a feasible edge fining algorithm is proposed 
to deal with the initial segmentation image. First, the boundary 
sub-blocks are distinguished by their classes and locations. 
Second, each boundary sub-block images are subdivided into 
four lower sub-blocks with size of (M/2) X (M/2). Texture 
feature of this level sub-block images are calculated. Finally, 
the distances of this level sub-block image to its adjacent 
classes are calculated. It will be marked as corresponding class 
according to distances. Actually, boundary sub-blocks should 
be altered to their neighbouring classes, which assured the 
integrality of the area of the segmentation. 
The algorithm is given as follows. 
Step 1: To divide the image I into sub-block images with size 
of M*M and calculate their texture features respectively. 
Step 2: To classify the sub-block images into certain number of 
classes using K-mean clustering algorithm. And then, pixels in 
image are marked to corresponding classes according to their 
texture feature. After that, the initial segmentation image S(0) 
would be gotten. 
Step 3: To distinguish the boundary blocks and divide them to 
lower sub-blocks with size of (M/2) X (M/2). Texture feature of 
this level sub-block images are calculated. These lower 
sub-blocks are marked to corresponding class according to their 
distances to the neighbouring classes. After that, the first fining 
segmentation S(l) will be finished. 
Step 4: If S(t)=S(t-l) or M/2<size of the smallest texture 
foundational element, then output the final segmentation image. 
Otherwise go to Step 3. 
4. EXPERIMENT RESULT 
4.1 Experiment 
First, the camera image is used for segmentation experiment. 
The original image and the segmentation image are given in the 
follows. 
4.2 Resu 
Figure 1. Original camera image 
imm 
■: *■ j 
t ■ • ■ - 
wf • * . 
.. L .1 
Figure 2. Segmentation image 
Second, high spatial resolution RS image acquired by 
unmanned aerial vehicle is given as the original image. The 
figure 3 is the original image and the figure 4 is the 
segmentation result. 
Figure 3. Original high resolution RS image 
Figure 4. Final segmentation 
292 
From th< 
segments 
However 
small sc r 
whole c 
that are 
The first 
Especiall 
value for 
The sec< 
clustering 
segmenta 
to give ' 
The thirc 
features 
the mean 
produce 
However 
is hard to 
By use ol 
can oven 
methods 
images, 
spatial re 
REF 
LIU L.F 
Sensing 
18(6):441 
Ming D 
algorithm 
engineeri 
HUANG 
Spatial R 
Transforr 
Universit 
Su J.Y. 
Residenti 
Geomatic 
29(9):791 
Zhou F., 
Peking U