163 
if | Ax, |> Tx or | Ay, \> T y , (Ax, A y )-(Bx, By) is 
mismatching points. 
6. EXPERIMENT RESULT 
The data for experiment are divided into 4 groups. 
In the first group, A1 is TM image in band 3 whose resolution 
is 30m and size is 726 X 623. B1 is SPOT image whose 
resolution is 10m and size is 1360 X 1138. Image A1 is obtained 
in August 2000 and B1 is obtained in May 2000. 
In the second group, A2 is TM image in band 8 whose 
resolution is 15m and size is 185 X 1105. B2 is SPOT image 
whose resolution is 10m and size is 1360 X 1138. Image A2 is 
obtained in June 2003 and B2 is obtained in May 2000. 
In the third group, A3 is aerial image whose resolution is 10m 
and size is 1024 X 1024. B3 is SPOT image whose resolution is 
5m and size is 2048 X 2800. Image A3 is obtained in October 
1993 and B3 is obtained in September 2004. 
In the fourth group, A4 is IKONOS image whose resolution is 
lm and size is 8831 X 9090. B4 is SPOT image whose 
resolution is 10m and size is 1360 X 1138. Image A4 is obtained 
in October 2003 and B4 is obtained in May 2000. 
The four group images and registration results are shown as 
below. 
There are much difference between A1 and B1 on acquisition 
time, ground resolution, sensor and gray feature, moreover, the 
rotation angle is 5 degree. 
Fig. 1(A1), (Bl) show the result of registration for TM Image 
and SPOT image (Threshold J' q for Forstner is 0.6), the feature 
points distributes evenly because of the evenly control. Under 
the initial steering by 6 pairs of artificial selection exact 
homologous, we get 137 pairs of homologous using invariant 
moments similarity measurement combined with global 
relaxation matching. Finally, the points whose number are 59, 
97, 134 and 135 is eliminated because they are mismatching 
points by using of the method of quadratic polynomial model 
detection. 
We used other 2 kinds of method to process the four groups 
data and the result is given by table 1. Method 1 is proposed in 
document [4], Method 2 is proposed in document [6], Method 3 
is proposed in this paper. 
From the experiment results, and compared with other method, 
we found this method has the suitable matching precision, the 
matching error maintains about 0.5 pixels. The matching 
method proposed in the paper, adapting the multi-sensor remote 
sensing image with high accuracy, in homogenizing control of 
feature points, mismatches elimination as well as the global 
relaxation algorithm, has the optimal result in the matching 
speed and the precision. 
(Al) 
(B2) 
(Bl) 
(B3) (B4) 
Figure 1. The result of registering between multi-source remote sensing images