The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B4. Beijing 2008 
1366 
3.1 Character Points Extraction 
Figure 1, Basic Principles 
3. STEREO MODEL AUTOMATIC 
RECONSTRUCTION FOR HIGH RESOLUTION 
IMAGES 
To realize the auto reconstruction of stereo model, the main 
problem is to resolve the auto extraction of the corresponding 
image points on stereo image pairs, which is the key step to 
realize stereo auto reconstruction and is one of research content 
of computer vision. A new technique to realize stereo character 
auto extraction and high precision orientation is provided 
according to the character of the high resolution remote sensing 
images. 
Character extraction is the basis of the image analysis and 
image matching and it is also the most important task of single 
image processing. The character in the digital images can be 
divided into point characters, linear characters and area 
characters. Point characters refer to those distinct points such as 
acute point, circle points etc. The arithmetic operators used to 
extract point character is called interest operators or favorable 
operators. The goal of image processing is to extract as more as 
possible and precise orientation point characters from images. 
At present, many interest operators have been brought out and 
they can be divided into three classes according to their 
property. The first class is based on shape, that is to say the 
character point is positioned at location with the biggest 
curvature of the outline or at the point of intersection of two 
lines. The second class is based on signal. The digital image 
uses discrete grid to record continues information so that the 
image processing is one of application area of signal processing 
and the method used in signal processing can also be used into 
image information processing. The third class is based on 
template, but this kind of operator is designed for idiographic 
point character. Also the accuracy of template operator can 
achieve sub-pixel but it is inapplicable and its main application 
area is close photogrammetry. 
Stereo remote 
sensing image 
pairs 
7 
I he least square 
mistake 
imitation of the 
Harris operator is a signal based operator for point extraction 
provided by C.Harris and M.J.Stephens in 1988 [1]. Harris has 
such characters: calculate simply, the acute point extracted 
distributed symmetrically, the characters can be extracted 
quantificationally and the operator has characteristic of stability. 
The operator is illuminated by the self correlation function used 
in information processing and given a matrix M which relates to 
self correlation function. The eigenvalue of the M is the first 
factorial curvature of the self correlation function. The point is 
considered to be a character point if the both curvature is high. 
The process is described as following: 
M = G(s) 
g x 
g x g 
g x g 
g\ 
I = det(M ) - к -tr 2 (M ),k = 0.04 
(1) 
(2) 
Where: 
^ x is the grads of the x direction; ^ y is the grads of the x 
direction; j s G auss template; det is the DET of the 
matrix; tr is the trace of the matrix; ^ is the default constant. 
The value of each element in the matrix I corresponds to the 
interest value of the initial image. The detection results using 
Harris operator refer to Figure 3. The corresponding acute 
points on the high resolution stereo images can be detected 
effectively as shown in the Figure. 
Figure 2, Technique flow of stereo model automatic 
reconstruction for high resolution images and the space 
information extraction 
Figure 3, Stereo image and Harris operator extract results