161 
points using Forstner operator, we extracted some initial points 
using Roberts operator[ 10] [ 11 ] „ 
2.1 Extraction algorithm 
d 1 = 
1 £ c, r 
~ £ c + 1, r 
d 2 = 
1 8 c, r 
- £ c , r + 1 
d 3 = 
\ & c, r 
& c — 1, r 
d 4 = 
\ & c, r 
~ 8 c , r - l 
1) Initial points extraction 
Computing the image difference in 4-neighborhood using 
Roberts’s operator is equal to calculating the four gray- 
difference absolute value d l ,d 2 ,d 3 ,d 4 : 
(1) 
Giving threshold T, The point ( C , r ) will be regarded as a 
initial point ifM — mid{d\, di, d3,d4} > T. 
The efficiency of extraction is influenced by T. Low threshold 
will increase the computational amount and high one often 
leads to omission and false point. Generally, the threshold is set 
to be 60 percent of the mean of the difference image. But this 
value should be adjusted according the gray and geometry 
feature of the remote sensing images. If the feature points 
extraction is too slow, the threshold should be increased, 
otherwise it should be decreased. 
2) Accurate extraction using Forstner operator 
The covariance matrix N and roundness q c of error ellipse 
in the 3X3 window around initial point ( c, r ) can be 
computed according to equation (2). 
(2) 
N = 
i*,g, _ 4 Det N 
I g.g, I g\ ’ q ' r (trN) 2 
Where ( C , r ) is the center of the window, g x ,g v are the 
difference in x and y direction respectively, DetN is the 
determinant of N and trN is the trace of No 
For threshold J' q , if q c r > T q > ( c > r ) is a choice point, 
then we compute the weight value Wc,r = DetN/tfN ^nd 
select the extreme points in the grid as the feature points 
according to ^ c,r . And ^ q is a empirical value, generally we 
T 
set q to the range form 0.45 to 0.7[8] [14]. 
2.2 Uniform Control 
In the process of feature points extracting, we must distribute 
the feature points evenly in the image rather than cluster in 
some local region. In the paper, we adopted Grid Control 
Technology base on entropy to ensure points’ distribution 
evenly. 
1) Dividing the image into small grid block evenly, and then 
computing the entropy for each block. The entropy is the 
measurement of the image information quantity and feature 
existence, the information quantity of image with lots of feature 
is more than the one with little feature [5] [7]. 
k 
E = -Y J Pm l °S(P m ) № 
m=0 
Where k=255, p m is the probability of the pixel with gray m in 
the grid block. 
2) These entropy value Eij of grid block are listed in order 
from large to small, and dividing all the block into 3 levels 
according to the grid size from big to smalk the quantity ratio 
of grid block is 2 :1: 1. 
3) In the first level grid block, extracting the feature directly 
because of the large contents of information and abundant 
feature. The feature points maybe too concentrate because of 
the non-uniform distribution in the first level part. But the 
feature points need to be evenly distribution in the whole image 
for the registration precision. 
4) If the feature points of the first level grid block is not evenly 
distribution, we extract the feature points in the prior grid block 
in order from the second level one, which make sure that there 
are enough feature points in the abundant information region, 
and there are also a few feature points within the less 
information region. 
3. SIMILARITY MEASUREMENT BASED ON THE 
VARIANT MOMENTS 
Given image f (X,y ) , the ( p + q ) level moment is 
— = IZ*' y q f (x, y ) ■ The (p + q) level central 
* y 
moment is U pq = X ~ Xo) P (y - yof f(x,y) , 
* y 
where Xo = TYl\o/Wloo, yo = YYlwj 17loo is the coordinate for the 
gravity center of the image» 
Normalizing all the moments using 0 level central moments, 
we can get the normalized center moments. 
T| pq — Upq/U 00 
where r = ( p + q)/2, p + q = 2,3 (4) 
Using the 2, 3 level moments, we can construct 7 invariant 
moments [3] [9] which have the invariability of translation, 
rotation and proportion. 
We can construct feature description vector for image using 
these 7 invariant moments. The euclidean distance for two 64 X 
64 image blocks whose center points are A and B respectively 
could computed as below. 
The similarity measurement is p(A,B) = » where 
P €E [0,1] and R is a constant.