The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
469 
Figure 1 shows the results of edge extractions through wavelet 
transformations in one channel but with different scales. 
From this we can see that by using large scales some details in 
the image are neglected and the main edge features are 
intensified; on the other hand, with smaller scales more details 
are kept. 
Generally, in buildings’ reconstructions we are focusing more 
on their outlines of images, and their details can be enriched 
with textures. We can efficiently extract edges of images with 
appropriate scales according to actual needs. We can also 
construct multi-scale edge images in the same channel with 
the use of the anti-symmetric multi-scale wavelet filters, thus 
facilitating the restriction of noise (Cheng, 1998). 
The wavelet filter proposed in this paper is based on edge 
extremum. Under the induction of maximum and direct map, 
it can locate the edges at sub-pixel accuracy, and doesn’t 
generate fake edges. So it is superior to the method of zero 
cross method (Peng, 1999). 
2.3 Edge thinning, tracking and parameters representation 
Where, x it y, = the node coordinates of feature vector and 
A/, = yl(x-x i _ ] ) 2 +(y i -y i _ l ) 2 
cos(c i x 45°) c, is even number 
cos(c ; x 45°) c, is odd number 
sin(c,. x 45°) cj is even number 
\/2sin(c,) c, is odd number 
At, = 
Ay,- 
there are five center moments can constitute the directional 
invariant moments for later feature matching. 
3. AUTOMATIC CORNER POINT RECOGNITION 
It is easy to get the homologous points via image matching 
using the comer points or inflection points of the building. 
Then we can obtain the contour feature. After analyzing the 
SUSAN algorithm (Smith, 1997), a new method of comer point 
recognition based on wavelet edge detection and linear 
template was proposed. The method could be summarized as 
follows: 
The result of edge detection is to get edge lines and Y-tracks of 
image structure, which have cryptic information of knowledge 
of the image. We can get the knowledge of edge feature via 
edge detection, thinning, tracking and parameters’ 
representation. So we can obtain full understanding about the 
information of knowledge of image from a lower-grade to a 
higher-grade level. With this knowledge we can get the shape 
of the object in 3-D space. 
We describe edge features by the use of orientation chain 
coding because of the obviously structured feature of building 
in large scale remote sensing images. The orientation chain 
code has a simple structure and powerful ability to describe the 
direction and is available to further representation of the line 
moments for the purpose of feature parameter matching. 
Here we improved the representation of an 8-neighborhood 
chain code. The improved chain code can indicate the 
dimensional direction of the feature vector and provide the 
important constraint condition for feature matching (Chen, 
2003). 
The line moment can represent edge feature very well because 
of its invariability for translation, scale and rotation. In addition, 
it has good computational efficiency. Moreover, we can easily 
get the moment characteristic quantity of any section of the 
edge, which is useful for comparing the local similarity. The 
equation (10) gives the low-level center moments of line 
features. 
h,o = Z Al i x f = Z A/ < + \ tei J 
Mo,i = Z = £ A/, j^ M +1 Ay,, j 
h.i=Z A to = Z A/ /[ x M + ^ Ax <]^ + ^ A T,J > 
2 
M-2,0 = Z A h x i — Z A ^< ^ x i-\~2 A *' j 
Mo,2 = Z'A/,-T, 2 = Z A/, (tt-j Ay, j 
(10) 
1) To determine the approximate position of the comer point 
on the image and intercept the window’s image centering on 
the point, feature the extraction to get the edge feature image of 
the window, which is called image aggregation A according to 
the view of morphology (Chen, 2003). 
2) To set a round template whose radius is R, the radius 
could be determined by the rough offset from the centre of 
windows to the exact position of the comer point. In figure 2, R 
is 3 pixels, the values of 8 direction elements of the template 
are non-zero, and the remained elements are zero. The template 
is called structural element B, which is also an image 
aggregation and suitable to the structural feature of building. 
3) To move the template in the window image, and cap 
image set A by using the structuring element B. Then get a new 
set C, 
C = AnB (11) 
In fact this step is to probe the image set by using a structural 
element. The elements Cj(x,y) are non-zero in the set C only 
when the edges are exist in the probing windows, 
c i (x,y) = c i (r)* 0 (12) 
and then count the non-zero element in a new image set to get 
the number of feature points in every probed windows, 
Yi Lin, Ph.D, Associate professor, Department of Surveying and Geoinformatics, Tongji University.