The International Archives of the Photogramme try, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
detected by using loose thresholds, but gained lots of false 
edges caused by interruption and noise, as shown in Figure 6(c). 
Using the proposed algorithm, a lot of weak edges of runway 
has been detected. Especially when comparing dotted elliptic 
area in Figure 6(c) with the area in Figure 6(d), the proposed 
algorithm can detect the runway edges which are missed by 
Edison algorithm using loose thresholds. But a part of false 
edges couldn’t be deleted, such as the dotted rectangular area. 
In city-modeling, the roof structure presented in ridges is 
important information of building reconstruction. The ridges 
belong to the weak edges because of minor intensity difference 
around them. The image of second experiment was taken from 
residential area, as shown in Figure 7(a). A few ridges could be 
detected by Edison using tight thresholds, as shown in Figure 
7(b), more ridges could be detected after using loose thresholds, 
but some false edges also were detected, as shown in Figure 
7(c). Using proposed method, the relative completed ridges has 
been detected, and has inhibited the interruption from 
surroundings, especially the false edges brought by trees in the 
dotted rectangular area, as shown in Figure 7(d). Unfortunately, 
some line segments on the road haven’t been removed. It is 
because the linear edges in the roads follow the rule of weak 
edges. 
The experiments compared proposed algorithm with Edison 
algorithm using two aerial images of different areas. The first 
experiment could certify the ability of detecting weak edges for 
less interruption included in the image; while the second 
experiment employed an image with more interruption to 
certify that it can inhibit false edges. Because tight and loose 
thresholds are necessary to the proposed method, so the 
procedure of experiment is that: the both tight and loose 
thresholds were applied to Edison algorithm respectively, and 
two edge detection results were gained at first; then the 
proposed algorithm used the both thresholds simultaneously to 
detect edge and gained another result; compared the edge 
detection results at last. The key parameters in the experiments 
are shown in table 1. Where Rank is normalized ranking of 
gradient magnitude. The image adopted by the first experiment 
is taken from an airport, as Figure 6(a), including airplane, 
runway, few buildings and grass. Edison algorithm detected 
some edges by using tight thresholds, such as the boundary of 
buildings and figure of airplane, but the boundary of runway 
was missed, as shown in Figure 6(b). The runway could be 
Windo 
w size 
Hysteresis high threshold 
Hysteresis low threshold 
Rank 
Confidenc 
e 
decision boundaries 
Rank 
Confidence 
decision 
boundaries 
Tight threshold 
5X5 
0.93 
0.96 
Box 
0.99 
0.91 
Ellipse 
Loose 
threshold 
5X5 
0.93 
0.93 
Box 
0.77 
0.77 
Ellipse 
Table 1. The key parameters in the experiments 
Results of the first experiment, (a) Input Image, (b) Result of Edison 
algorithm with tight thresholds, (c) Result of Edison algorithm with 
loose thresholds, (d) Result of the proposed algorithm.