412 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
precisely obtained. The whole extraction flow is shown in 
Figure. 1. 
Most of these surveyed methodologies on building extraction 
from monocular images can be classified into two categories: 
edge-driven and region-driven. An edge-driven approach uses 
an edge map as its starting point usually followed by a line 
extraction process in order to reduce the numerous spurious and 
insignificant edges that are found. However the lines including 
the building-correspondence lines and some irrelevant lines are 
fragmented and distributed randomly in the scene. A mass of 
endeavours (V. Venkateswar and R. Chellapa, 1986; A. Huertas 
and R. Nevada 1988; R. Mohan and R. Nevatia, 1989; R. Irvin 
and D. McKeown 1989; Y. Liow and T. Pavlidis, 1990) have 
been made to link or group the line segments corresponding to 
the building to obtain desired building boundaries. This is the 
primary difficult of the edge-driven approach. And usually it 
needs a complex bottom-up grouping process. In a region- 
driven building extraction strategy, the source image is initially 
segmented entirely into different regions. After segmentation, 
an attempt is made to determine which regions are 
corresponding to the building component and to combine these 
building related sub-regions because the exact building region 
may be segmented into many sub-regions. At present, some 
task-specified automatic approaches or semi-automatic 
approaches to address this issue are applicable which are 
implemented by the manual selection of the building sub- 
regions to form the building outline. This approach avoids the 
complex process of bottom-up shape recognition and formation 
in the edge-based approach in certain extent. The schema 
supposed in this paper is fall into this kinds of strategy. 
This paper is organized as follows: we begin with an overview 
of our approach in section 2, while in the following section 3 a 
briefly demonstration of seed region grow algorithm is given. In 
section 4, we will elaborate the two schema devised to form the 
precise shape of the building. In the last section, some 
extraction result and a discussion will be presented. 
2. METHOD OVERVIEW 
This schema is a region-based approach. Seeded region grow 
algorithm (Mat-Isa. N. A, 2005) is first applied to collect pixels 
inside building regions to form approximate shape of building. 
In order to extract the regular building boundary, two schemas 
are supposed which can be applied in different scenes. When 
the building polygon can be represented as a convex polygon, 
we can calculate the convex hull of the building from the pixels 
collected in the growing process. When the building polygon is 
not convex, the boundary of the building can be obtained by 
boundary fitting on the condition that the contrast between 
building and background is large enough. If the contrast is low, 
Hough transform is applied to the region image derived after 
the growing process. The dominate line sets of the building, 
which is perpendicular to another, can be retrieved. The 
intersections of two lines set construct a node matrix. If there 
are m X n lines in two dominate orientations, the node matrix 
is m by n. Based on the classification of orthogonal comers; 
shape of buildings can be represented by a tag sequence which 
is a series of symbols of related right angle comers. Based on 
the assumption that building is comprised of orthogonal comers, 
a filtering of orthogonal comers can be carried out to eliminate 
some false comers. The fine extraction process is implemented 
via matrix search algorithm. The building boundary can be 
Figure. 1 Extraction flow 
3. SEED REGION GROW ALGORITHM 
In this algorithm, the user needs to determine the region of 
building by manually selecting the position of the building in 
the image after filter. And the user also needs to define a 
threshold. 
3.1 Algorithm 
1) 
2) 
3) 
4) 
5) 
Manually select the seed points of building region. 
Chose NxN neighbourhood of the seed points. Calculate 
the mean value x and the standard deviation (J of the 
NxN neighbourhood. 
Grow the seed points to its neighbour’s pixels. Compare 
the grey level of the seed points with its neighbour pixel. 
Include the pixel into the region if it satisfy one of the 
conditions listed below: 
a) If the gradient of the pixels is less than 95% of the 
equalized histogram and the grey level of the pixels 
is less or equal to the predefined threshold. 
b) If the gradient of the pixels is more than or equal to 
95% of the equalized histogram and the grey level 
of the pixel is not more than or equal to one 
standard deviation away from the region mean. 
Set the neighbour pixel of the seed point, which is added 
to the region in the previous step. 
Repeat step 2 to 4 until all the pixels have been 
considered to be grown or the pixels cannot be grown 
anymore. 
4. FINE EXTRACTION SCHEMA 
After the steps above, an approximate region of the building has 
been derived. In order to extract the precise and regular 
boundary of the building, the following two schemas are 
supposed.