2004 
  
  
ng 
cause 
erent 
if we 
ction 
ween 
1 low 
hway 
most 
es of 
who 
on is 
And 
E L3] 
dr 4 
cem 
  
| Ex 
near, 
ition 
erent 
o In 
ature 
ion. 
ects 
ature 
in be 
tures 
most 
ther 
parse 
>]s is 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
according to the difficulty of extracting the features and 
complexity of the correlating computation for searching those 
features in the input image. The easier and simpler one is used 
in the lower levels. We think the last level template is the object 
image template. Here we have a example for explaining the 
establishing process of feature template for a simple object, 
ground vehicles in remote sensing image. We study vehicle 
image in several remote sensing image and we get the primary 
feature of vehicles. 
The profile of vehicle reflects the main feature of ground 
vehicle especially when we see ground vehicle from space or 
air. The background behind the vehicles also provide strong 
information for discriminate the vehicle. So we can think an 
object is a vehicle since it is on the road. Hierarchical vehicle 
templates can be formed as figure 4. 
  
    
     
  
  
Profile of vehicles 
Almost oblong shape 
  
Level 1 
  
  
  
  
  
         
  
   
    
  
en bor bac 
Road, parking lot, 
   
  
2 
Level 3 
[Level 4 
Figure 4. The hierarchical feature template for vehicles in 
remote sensing image 
    
  
The inherent feature 
Size, windshield 
  
  
  
Particular feature 
Sign, color, detail shape 
  
  
  
  
  
  
  
We use the profile of the vehicles as the first level template, the 
width of most of the vehicles is near 1.5-2meter, but the length 
is quite different, so we form several oblong shape templates 
with different ratio of width to length. Other level templates can 
be described by extracting corresponding features. In this paper, 
we focus on the first level template especially on the shape 
profile feature of the object, because the profile features are the 
most common features of objects. 
S. CORRELATION PROCESSING 
What we focus here is the template correlation based on the 
object shape template (most common used feature in object 
recognition). Because of the reason mentioned in the first 
sections, object recognition based on GHT(Ballard 1981) or 
Hausdorff distance (Rucklidge,1997) require rigid model 
representing the object to be found and do not meet the 
demands: robust to occlusions, clutter, arbitrary illumination 
changes and sensor noise, but in remote sensing image the 
assumption of a rigid model is not fulfilled the correlation 
match based on the distance between profile to shape center 
(DPC) is presented for recognizing the object with protruding 
profile in this paper. 
3.1 The definition of DPC of object image 
When we acquire the profile of the object image, firstly the 
location of shape center of the profile is computed. Assume 
that the profile point set is T(x,y;) I=1,2,3....N, where N is 
the number of points in set T. the shape center C(x,y) can be 
computed by equation C1): 
(1) 
N N 
Xs(VY x)/N Fed a N 
i=] i=l 
Then we can get the distant set S of DPC 
889 
  
$ü)9 Aix, - xy «(y,- yy 1=123...N (2) 
Where — (x,y;) 7 the coordinate of profile points 
(x, y) = the coordinate of shape center 
Normalized S(i)=S(i)/S 
max» 
  
y) (Ye «^ S) 
    
(XN. yx) 
(X Yi) 
  
  
  
Figure 5. The DPC set of object template 
The DPC set reflects the distance change process of the profile 
related to the shaper center. So there are two advantages when 
the DPC set is used as the feature for correlation process: 1) this 
processing has the characteristic of translation invariant 
because the shape center is used as a reference points; 2) this 
processing has the characteristic of rotation invariant due to the 
method for computing the DPC. 
5.2 Correlation processing 
DPC convert the two dimension edge profile to one dimension 
feature, it is convenient to choose a correlation method for 
template matching. The DPC mean square difference (MSD) 
between the object profile acquired from input image and 
template is proposed as template correlation criterion. And a 
threshold is set to decide object image is similar to template or 
not. Assume that S,(n) is DPC set of object in input image and 
Sm (n) is the DPC set of template, n=1,2,3...M, if N<M, S,(n) 
will be expanded to M by interpolating processing. Here we 
think N=M. the correlative conjunctions based on MSD can be 
written as equation (3): 
N 
R(n)- Y GO - 8, G0 0) 3) 
R = min(R(n)/ N} 
Where  R (n) is the correlative conjunctions 
R is the correlative degree between object and 
template 
> <> 6 Accept (4) 
Ng Refuse 
Where o is the threshold for making decision that the object 
is similar to template or not. 
From this process we can see that the rotation direction 
difference between object image and template does not affect 
the result of correlation. Table 1 list the simulation correlation 
results of multi basic shape, where row direction are the 
template, and column parts are simulated objects which have 
some change with template. The number parts represent the 
correlation degree. 
From the table 1 we can see that correlation based on DPC set 
has good discrimination to basic shape, at the same time we can