hould be 
bviously 
| method 
use from 
. method 
 artifical 
0 contin- 
satisfied 
d will be 
:cond di- 
require- 
rge then 
us point 
of corre- 
> original 
pixels), 
or exam- 
; thresh- 
than the 
| image, 
ble, if it 
1 results 
ed flexi- 
f image, 
edge ex- 
ng image 
on, dis- 
ic search 
atic con- 
f A* al- 
ium path 
ost of an 
le. 
| started 
in of dis- 
on are: 
discrete 
hese dis- 
crete points will have the possibility to be connected in- 
to line, otherwise it can not be connected. 
(2). In the selection of equidistance discrete 
points, cost function should be calculated according to 
the following formula and the pointwith minimum cost 
will be considered as connection point. 
cost = G, — G, + 2 (GD, — GD) 
where, G,, G, is the grey value of point n and object 
point 0, GD,, GD, is the first derivatives of 
point n and object point 0. 
(3). When the number of automatic looking for 
connection point is small then threshold , then the edge 
will be removed automatically. 
(4). The searched edge points will not be searched 
again so as to ensure edge point has the limit of search 
once. 
(5). It is prohibited to search in opposite direction 
in the automatic search so as to avoid mistaken connec- 
tion. 
3. Manual Editing 
After automatic connection, the discrete points 
which can be connected all have been connected into 
curve, however, as a result of the edge of remotesens- 
ing image is very complicated, part of defects still might 
exist even if many measures have been taken, so it must 
be through manual editing. 
After automatic connection and editing, the image 
boundary is formed and it is very useful for practical to 
use. 
4. Edge Extraction Test and Results 
The test image is TM and obtained in July 1987, 
located in Ji county of Hebei Province. Band 3 is select- 
ed and the image size is 512X512 pixel. Medium filter- 
ing is firstly used with 5X5 window to eliminate the af- 
fects of noise, then first and second derivatives are de- 
termined also using 5 X 5window, the point of which 
first derivatives is larger than 5 and second derivatives 
has zero crossing is taken as edge point. In terms of 
thestatistics, first derivatives are from 6 to 65. It can be 
seen that the change range for first derivatives of corre- 
sponding edge point is more wide, so it is not suitable to 
use a threshold to control an image. Image should be di- 
vided into subimage with 110 X 110 pixel size and the 
threshold of each subimage should be adjusted by visual 
monitor method. After the adjustment, the threshold 
change of each subimage is from 8—25. 
Then initial editing will start, this step is mainly to 
delete evident mistake. 
After the initial editing, discrete point will be con- 
nected using automatical method. The maximum search 
45 
region is 11 X 11 pixel. Then the final editing will be 
conducted, through deleting, addition and modification, 
the results of image edge extraction can come upto the 
optimum standard and is shown in Fig 1. From the pro- 
cessing results it can be seen that image has been divid- 
ed into reasonable blockette and reached the goal of im- 
age segmentation. 
In order to for comparison, Gauss — Laplace 
method has been used, the result is shown in Fig 2, in 
the test, window size is 11 X 11 pixel, standard devia- 
tion is 1. 4, threshold is 5. From the the comparison, 
the new method introduced in this paper is better than 
Gauss — Laplace operator. 
5. Conclusion 
Research on edge extraction has been conducted for 
many years, the edge feature points can be extracted for 
many methods. These feature points can be used for im- 
age matching. but there are some difficulties in image 
analysis. For the requirement of image analysis in the 
feature, it isbetter to provide continuous boundary. For 
this reason, a research of boundary extraction based on 
zero crossing of second directional derivatives has been 
carried out. 
This paper describes the principles of edge extrac- 
tion using zero crossing of second directional deriva- 
tives, the formula for fitting of two dimension image, 
first and second derivatives have been derived. In order 
to overcome the difficulties of dynamic threshold in edge 
extraction, threshold will be determined using manual 
interaction. For reasons of edge extraction to be contin- 
ued, curve tracking can be conducted using artificial in- 
telligence search method and discrete point will be con- 
nected automatically, and it has flexible editing func- 
tion, which is able to correct the defects in edge extrac- 
tion conveniently. The test proves that the edge ex- 
tracted by this method is better than Gauss — Laplace 
operator. 
REFERENCE 
1. Lu Yan, A New Algorithm of One Dimension 
Edge Extraction using Sequence Analysis, sym- 
posium of Digital Photogrammetry, Wuhan, 
China, 1988. 
2. L. Vliet & I. young, A Non — Linear Laplace 
Operator as Edge Detector in Noisy Image, 
Computer Vision, Graphics and Image Process- 
ing, 1989. 
3. Zheng Zhaobao, Image Matching Method via 
Dynamic Programming, Acta Geodetica et Car- 
tographica Sinica, Vol 28, No. 2, 1989.