The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B3b. Beijing 2008 
510 
highways. This paper is organized as follows. Section 2 will 
present the question of “what is phase classification”. Section 3 
will explain the thesis of this paper used in the work of road 
extraction. Section 4 will show the experiment on the road work. 
The final section will give the conclusion to the improved 
approach on road extraction in high-resolution remote sensing 
image. 
2. PHASE CLASSIFICATION METHOD(RAO H., 2004) 
General algorithms of edge detection often refer to the gradient 
information of image, such as gradient algorithm, Laplacian 
algorithm, Marr algorithm and so on. The main idea of it is to 
calculate the gradient scope of one point whether it is over the 
pre-threshold. If over this pre-threshold, there will exist the 
edge, or there will not. Base on the method of gradient, it is 
thought that the edge exists in the changing area of the grey. 
But the effect is not obviously to the images with even grey. 
Bums and other persons thought the edge exists not only in the 
changing area of grey, but also in the place where the grey 
changes in one following direction. So Bums proposed a 
method for the work of extracting the straight line based on the 
phase characters of edge, which is named phase classification. 
This method is very effective to extract the edge and outlines of 
images. The exact algorithm in the following is: 
Figure 1. Chart of Phase Classification 
Input the pixels of one image line by line, and then calculate 
each difference of every edge pixel in the direction X, Y, which 
is D x and D v . Finally we can get the tangent value of each 
point, arctg = (D v /D x ) • It is also the gradient direction, 
which refers to the most severe direction compared to its 
adjacent points. The relevant formulas are: 
D x = p[x -2,y + \} + p[x -\,y + \]y. 2 + p[x, y + 1] 
-p[x - 2,y -\\ - p[x -\,y -\]*2 - p[x,y -\] 
(1) 
D y = p[x,y-\]+ p[x,y]*2 + p[x,y + \]-p[x-2,y-\\ 
-p\_x - 2,y]x 2 - p[x - 2,y +1] 
(2) 
6 = arctg(D y /D x ) (3) 
M = \D x \ + \D y \ (4) 
where D x = horizontal difference. 
D v = vertical difference. 
0 = gradient direction. 
M = gradient scope. 
In the method of phase classification, put all the points that 
belong to the same direction area and connect each other in 
geometry into one class, which is the edge class. In fact, the 
grey of most points in image changed gently, so we make the 
value M over the pre-threshold into the classification. In order 
to reduce the processing data column in the following steps, we 
see the adjacent points, as well as have the same value 0, in the 
horizontal direction as one code. Then compare the code of one 
row to the one of the previous row, class the points which have 
the same code into the same area. After doing these works, 
mark all the edge points with the number one to eight. Like 
figure 2: 
90 
Figure 2 Chart of Dividing the Gradient Directions 
After divide all the edge point, fit them that belong the same 
area with the method of least square to extract the edge line 
exactly. 
3. ROAD EXTRACTION BASED ON PHASE 
CLASSIFICATION 
Road in high-resolution remote sensing image includes 
following five characteristics: 
1 ) Narrow width and smooth curve; 
2) Direction changes smoothly too; 
3 ) Internal grey values are even; 
4 ) Gray values inner differ much from that of background; 
5 ) A certain length of the road; 
Aiming to these characteristics above, this paper proposes a 
new method in extracting road information from high-resolution 
remote image. This method is based on phase classification. 
Firstly, define threshold automatically taking advantage of 
canny operator. Secondly, transform the original image into a 
bi-value one. Finally, improve the phase classification method 
in order to extract different kinds of road. This approach is 
almost automatic without given initial coordinates as well as