The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
646 
registration, under the condition that the images have 
preliminary registered. 
2. IMAGE REGISTRATION MODEL 
2.1 Principle: Collinearity of Corresponding straight lines 
Straight lines are familiar features in airborne image sequences. 
They often fall into edges of man-made objects. Since edge 
extraction is one of well-studied techniques in image processing, 
we can design a robust algorithm of automatic straight line 
extraction from edges, with which straight line will be extracted 
from source image and reference image respectively. Then the 
extracted straight line is acted as control object for the 
registration process. 
As shown in figure 1, we suppose that the source image has 
rotational and translational distortion compare to the reference 
image. Let If and If (/ = 1,2,3) represent three corresponding 
straight lines in the reference image and Source image 
respectively. After rotating and translating, the source image is 
located at the right position, which is the position of the 
registered destination image. Here the straight line if 
(transformed from L] ) in destination image and the 
corresponding straight line If in reference image are located in 
the same one straight line. Considering the difference of pixel 
values between the reference image and source image, each 
endpoint of corresponding straight lines if. and Jf does not 
need to be the same coordinate. This is the principle of 
registration: collinearity of corresponding straight lines. 
Compared with point feature-based method, the advantage is 
that the straight line can be specified by arbitrary two points in 
it, but not its’ endpoints exactly. Meanwhile, because the 
straight line can be extracted automatically and precisely, this 
method has the potential promotion of both efficiency and 
precision. 
reference image source image to be registered 
rotation of source image translation of source image 
Figure 1. Principle of image registration with straight line-based approach 
2.2 Registration Model for Image Sequences 
To establish the registration model, we adopt direct image 
registration method. That is, for every pixel in the source image, 
the corresponding pixel coordinate is found in the registered 
destination image. The coordinate transform formula 
x d =f(x s ,y s ) y d =g(x\y s ) (1) 
is used to resample image gray levels (i.e. colors) to construct 
the registered destination image. Where (x d ,y d ) represents an 
image coordinate in the registered destination image, (x' 5 ,/ 5 ) 
represents a pixel grid in the source image. The core task of 
image registration is to determine the form and parameters of 
formula (1). 
As shown in figure 2, the straight line If in the reference image 
corresponds to the straight line If in the source image. In this 
model, the endpoints of two straight lines need not to 
correspond one-to-one. The straight line L cl in the registered 
destination image is created from If after be registered. Its’ 
two endpoints are transformed from two endpoints of If by 
formula (1). According to the collinearity principle, straight 
lines if and If are in one straight line. In another words, every 
points in if are located in If . 
reference image registered destination image 
Figure 2. Model of image registration with straight line-based 
approach 
Let symbols p r , p r represents arbitrary two points in If, p d , 
p d represents arbitrary two points in if , respectively. In 
practice these points can be endpoints of If and if . Let 
symbol a represents the vector from p r to p' 2 , b represents 
the vector from p r x to p d , c represents the vector from p d to 
p d , respectively. The expressions are: 
n = (x r 2 -x[)i + (y r 2 -y[Y] 
b = (xf -xOi + Gyf -y[)\ 
c = (x 2 -x d )i + (y d -y d )j